Rochester Institute of Technology Rochester Institute of Technology

RIT Scholar Works RIT Scholar Works

Theses

1987

Adaptive statistical recognition of hand-printed Telugu characters Adaptive statistical recognition of hand-printed Telugu characters

Murthy Lakshmana Mantha

Follow this and additional works at: https://scholarworks.rit.edu/theses

Recommended Citation Recommended Citation Mantha, Murthy Lakshmana, "Adaptive statistical recognition of hand-printed Telugu characters" (1987). Thesis. Rochester Institute of Technology. Accessed from

This Thesis is brought to you for free and open access by RIT Scholar Works. It has been accepted for inclusion in Theses by an authorized administrator of RIT Scholar Works. For more information, please contact ritscholarworks@rit.edu.

RO(H~ST~R [~STIT~TE OF TEC~NOLUGY SCHOOL OF COMPUT~K S(IE~Cf A~G TECHNCLGGY

ADAPTIVE STATISTICAL kECOG~ITIO~ OF HAND-PR[NTEO TELuGU CHARACTERS

BY

~URTHY L~KSH~A~A ~A~T~A

lnls thesis IS sUb~itteJ t.o Tre Faculty of tne School of LOin put e r ') c i en c e 3 n J T p c h no Icy yin p,; r t , a I t u I f i I I men t of tne requirements for the degree of Mdster of Science in Computer Science.

Ap;HOlidls:

John A. Biles Thesis Auvisor: Professor John A. biles

L. A. Coon Committee Member: Professor Ldrry Coon

Peter G. Anderson Committee Member: Professor Peter Anderson

ABS Ik ACT

A brief description of statistical and syntactic pattern

matching techniques is presented with in eirphasis on

statistical techniques. The characteristics cf the Telugu

script are described. A subset of lb characters* which are

both easy ana hard to recogrize* is selected for the

dictionary of s t a r1 d a r a characters. A weighted linear

difference oolynomial of features is used to recognize

Telu]u characters. The features were Fourier descriptors of

projection profiles and cross sections t a k e n in various

directions. Algorithms for obtaining the projection

orofiles cross sections and adaptive learning iTethod are

presented. The system was trained and tested with a set of

8 nana-written samples of each of 16 different Telugu

characters. More than 90% of the 123 samples were correctly

recognized by the system. Results of numerous trials

examining the different features and classification

techniques are discussed.

ACK.nCaLEOGE^EiNTS

1 wish to scKnonledge my thanks tc Jeff Peltz, Faculty

Member, IrraZng Science Cepartment* for maKing the cigital

video equipment available for this study; uave Hedvedeff- at

the ROSS Computing Center for helping ir>e with VAX/VCS

syster, , ano AI Biles* my thesis advisor i for his excellent

jui dance ana critical review of the manuscript. Last but

not least* I would like to thank my friend Dersy Song* whose

thtsis wor.< in Chinese character recogintion has teen a

source of inspiration and reference to this study.

TABLE CF CC.\ TENTS

1.

1.1

1.2

1.3

INTRODUCTICN AND JVEkVItv,

INTRODUCTION TC PATTERN RECOGNITION

ABOUT TELUGU CHARACTERS

OBJECTIVE

Z THEORITICAL AND CONCEPTUAL DEVELOPMENT

2.1 TEMPLATE MATCHING

2.2 STATISTICAL METHODS IN PATTERN RECOGMTICN

2.2.1 FEATURE EXTRACTICN

2.2.1.1 METHCO OF VU:"ENTS

2.2.1.2 PROJECTION PMJFILZ

2.2.1.3 CROSS-SECTIONS

^.2.2 CLASSIFICATION METFCCS

2.2.3 LEARNING

2.3 SYNTACTIC METHODS IN PATTERN RECOGMTICN

2.3.1 LINGUISTIC PATTERN RECCGNITION SYSTEM

2.3.2 SELECTION CF PATTERN PMMITIVES

Z.i.i PATTERN GRAMMARS

2.3.4 SYNTAX ANALYSIS AS RECOGNITION PROCEDURE

2.4 DESCRIPTIVE CHARACTERISTICS OF TELUGU CHARACTERS

2.4.2 PREVIOUS *IGRK IN TELUGU CHARACTER RECOGNITION

2.4.1 RECOGNITION OF TELLGU CHARACTERS USING

STRUCTURAL MtTHOCS

3. RELEVANT THEORIES ANC CONCEPTS

j.2 t-EATURE ScL-CTIGn ANu EXTRACTION

3.2.1 PROJECTION PROFILES

3.2.2 CROSS-SECTICNS OR RUNS

3.2.3 CONDbNSED CROSS-SECTIONS

3.2.4 NOISE REMOVAL IN CCNLENSEO STRINGS

3.3 MAGNITUDE SPECTRA USING FOURIER ANALYSIS

3.4 ADAPTIVE LEARNING

3.5 CLASSIFICATION

4.

4.1

4.2

4.3

4.4

4.5

IMPLEMENTATION AND EXPERIMENTAL PROCEDURES

IMPLEMENTATION

HARDWARE CONFIGURATION

SOFTWARE CONFIGURATION

TEST LATA

TEST PROCEDURES

5.

5. 1

5.2

5.3

5. 4

5.5

6.

6.1

6.2

RESULTS AND DISCUSSION

RESULTS OF TEST 1

RESULTS OF TEST 2

RESULTS OF TEST 3

RESULTS OF TEST 4

RESULTS OF TEST 5

CONCLUSION AND FUTURE RESEARCH

CONCLUSION

FURTHER STUDY

7. HIBLlOCRAPhY

d. ACKNOWLEDGEMENTS

9. LIST OF ILLUSTRATIONS

LIST OF ILLUSTRATIONS

1.1 SAMPLE TELUGU PRINT

Zl STATISTICAL PATTEkN RECOGNITION SYSTEM

2.2 SYNTACTIC PATTERN RECOGNITION SYSTEM

2.3 CROSS-SECTIONS FOR A TAMIL CHARACTER

2.4 CROSS-SECTIONS FOR NUMERAlS 0-9

2.5 dLUCK LlAGkAM CF A PATTERN CLASSIFIER

2.6 PATTERN CLASSIFICATION dY PROXIMITY CONCEPT

2.7 P^IMITIVtS FOR A RECTANGLE

2.H HECTANGLE WITH PRIZTIvcS IN FIG. 2.7

2.9 TREE STRUCTURE FOR 2.7

2.10 STRUCTURAL DESCRIPTICN OF A PATTERN FRCM FIG. 2. 7

2.11 DIRECTED lA3EllE0 GRAPH FCiv TAMIL CHARACTER

2.12 SLOCK CIAGRAM OF LINGUISTIC PATTERN RECOGNITION

SYSTEM

2.13 BUILD PRIMITIVES FOR DEVANAGARI SCRIPT

2.14 COMPOSITION OF 3r3 ANC ITS PLANG DESCRIPTION

2.15 LABELLED PATTERN AND PARSE TREE FOR '^5'

2.16 FREEMAN CHAIN CCDE

2.17 PATTERN PRIMITIVES IN TERMS CF REGIONS

_.18CONTEXT SENSITIVE AND CONTEXT FkEE jRAM"ARS FOR A

rectangle

2.1-* representation of chromosomes

2.20 representation cf a house

2.21 head and tail primitives

2.22 PICTUkE DEFINITION LANGUAGE FOR A'HOUSE*

2.23A TELUGU VCrtELS

2.23B CONSONANTS

2.23C VOWEL SIGNS

2.23D EXAMPLES OF C-V LETTERS

l.lSt'

EXAMPLES OF C-V LETTERS

2.23F BASIC LETTERS

2.23G BUILD PRIMITIVES

l.lir\ CONJUNCT PRIMITIVES

2.231 CONJUNCT CONSONANTS hITH CONJUNCT PRIMITIVES

2.23J CONJUNCT CONSONANTS IN *HICH CCNSONANTS WITHOUT THE

FIkST VCwEL SIGN APPER ZLGW C-V LETTERS.

2.24 BLOCK DIAGRAM OF SYNTACTIC RECOGNITION SYSTEM FOR

TELUGU CHARACTERS

2.25 SAMPLE T-TUPLES

2.26 TELUGU BUILD PRIMITIVE AND ITS T-TUPLES

2.27 CODE IN TERMS OF T-TUPLES FOR TELUGU *y*

3.1 FUNCTIONAL dLOCK DIAGRAM

3.2 ALGORITHMS FGrt OBTAINING PROJECTION PROFILES

3.3 PROJECTION PROFILES OF CHARACTER '&

'

3.4 CROSS SECTION OF CHARACTER 'gr,'

4.1 HARDWARE CONFIGURATION

4.2 SOFTWARE CONFIGURATION

4.3 HAND-PRINTED STANDARD CHARACTERS

4.4 HAND-PRINTED SAMPLE CHARACTERS 1*2

4.5HAND- PRINTED SAMPLE CHARACTERS 3,4

4.o HAND-PR ImTED SAMPLE CHARACTERS 5,6

4.7 HAND-PRINTED SAMPLE CHARACTERS 7,3

4.6 HAND-PRINTED SAMPLE CHARACTERS 9,13

4.9 HAND-PRINTED SAMPLE CHARACTERS 11,12

4.10 HAND-PRINTED SAMPLE CHARACTERS 13,14

4.11 HAND-PRINTED SAMPLE CHARACTERS 15,16

4.12 DATA FILE STRUCTURES FCR STANDARD CHARACTERS

h.13 DATE FILE STRUCTRUES FCR SAMPLE CHARACTERS

5.1 SUMMARY OF TEST RESULTS

5.2 CROSS-SECTION VALUES AND FOURIER DESCRIPTORS OF

RRuJECTIGN PROFILES FOR Id STANDARD CHARACTERS

3.3 FUUKLIR DESCRIPTORS CF PROJECTION PROFILES AND

C'^O.S-SECTIONS FOR lfc STANDARD CHARACTERS

5.4 CONF US IIj,\ MATRIX FQk RESULTS OF TtST-1

5.5 CONFUSION MATkIX FOR RESULTS OF TEST-2

5.6 CONFUSI'JM MATRIX FCR RESULTS OF TEST-3

0.7 CONFUSION MATRIX (COLUMNS) FOR RESULTS OF TEST-4

AND 5

5.8 CONFUSION MATRIX (ROWS) FOR RESULTS OF TEST-4

AND 5

5.9 CONDENSED CROSS-SECTIONS FOR STANDARD AND SAMPLE

CHARACTERS

5.10 CONDENSED CROSS-SECTIONS AFTER NOISE REMOVAL FCR

STANDARD AND SAMPLE CHARACTERS

CHAPTER 1

INTRODUCTION AND OVERVIEW

1.1 Introcuct ion To Pattern Recognition.

The problem of pattern recognition has received the

attention of many researchers for the past 25 years. The

approaches taken to solve the prooleir of pattern recogr ition

fall into three general categories: (1) template matching,

(2) the classi f icatory approacn ana (3) the syntactic or

structural approach.

In the template matching method, an unknown pattern is

compared with the template of each pattern classes and the

classification is cased on preselected matching criteria.

In tne cl assi f icatory approach a set of characteristic

measurements, called 'features', is extracted from the

patterns, and the feature set is treated as a point in a

multidimensional space. The recognition of each pattern is

usually made by partitioning the feature space into

subspaces, each sufcspace corresponcing to a pattern class.

This partitioning* made by decision functions* m j :j h t he

effected cn tne basis cf statistical or non-statistical

considerations. This approach has oeen successfully usaC on

troblems like character recognition, medical diagnosis, crop

classification etc. The most important aspect of this niodel

TELUGU CnARACTcP r.-^CuM T I C N Pace 2

is the a priori selection of an optimum feature set and

decision functions. This can be a difficult task even for a

siitple pattern set. For large and complex sets like Chinese

idiographs and Telugu characters, it is not apparent how

this selection should be made.

In the syntactic approach, the central aim is to

generate the description of a given input in terms of its

subparts called primitives'. Pattern primitives are

selected a priori and tneir relations in the patterns are

descrioed by a grammar. The recognition process is

accomplished ty parsing the sentences cescribing the given

input patterns. Some of the problems associated with this

approach are: the selection of primitives is sirrilar in

nature to feature selection; the selection of a suitable

shape analyzer to recognize primitives is no easy task; and

the selection of a suitable grammar to describe the patterns

is difficult.

The key to pattern recognition does not lie wholly in

statistical approaches, heuristic prograrnr i ng , or irore

formal linguistic approaches alone. A good pattern

recognition system uses statistical, sturctural and

heuristic tools at various stages of processing of the

patterns, with e^ch tool being applied dt the stage to which

it seems oest suitea K a n 72J.

1.2 About T e I e u g u Characters.

TELUGU C H A >. A c T c !- R.Z l GM T I C n a:e 3

Telugu is one of the notional languages of Inoia,

spoken by about oC million people in the state of Anchra

Pradesh. The language h a _ a beautiful cursive script with

well over 2000 characters. Telugu is a highly phonetic

language with each character representing a syllable.

The Telugu alphacet consists of 16 vowels and 36

consonants, c onsonant-

vowe I combinations (C-V letters) and

conjunct consonants. A vowel following a consonant takes on

a _ i f f e r e r. t graphic form called a 'vowel sign', hence, the

vowels and consonants combine togetner to forir 5 76 different

C-V letters. The C-V letters ire for mad by addirg vowel

_ic,ns to consonants at appropriate t I a c e s . The C-V letters

ano consonants are come i red in many ways to procuce

thousands of different characters. Figure 1.1 shGws an

extract from a Telugu cook whicn illustrates some of the

commonly used styles and fonts.

The Telugu characters are not cf uniform dimensions.

However, they can oe conceived as different combinations of

special symbols, called primitives. The telugu typewriter

keyboard was designed under similar assumptions.

1.3 Objective.

Tne objective of this thesis was to investigate

classif icuory nethods for recognizing hand printec telucu

characters.

10vl

1^

hYJ

p

ja Ta

a S fl^

^ *.

C fe *3o ^ o

9 V- YJ

-j-J 13 Y3 Y2 k

V2.V3 O Y3 V3Q Xl -0

vl* t

0 Viv: jo

-J ta

*T 6

Vl fa

A fl * &

^ _._lo

-

Y2

V3

V$*

S1-3

O

9

^ 5 -a

"

* * a

3 YT?,1ys

YJ.J=

xjo>8

"l_

8*1o

YlL.

YJ

*

A O To fa

O to *j>

YlJ*

Yl _??

YJ

*uw gggBaaS-KSBB jB6_B_8R0BR__IjT3_>'1O_W_O__'

FlCS> 4 - i_ 5AlMPce TELCGrO PP|UT

T E l U G 0 C H i^'

A C T L-n t C _ GM T I C

'

Page 4

Two different methods of extracting features were used,

nairely: fourier descriptors of vertical, horizontal, left

and right diagonal projection profiles, and fourier

descriptors of row and column cross-sections.

Adaptive learning methods were used to train the system

with different samples to adjust the weights of features

according to recognition or uisrecogn ition.

Very encouraging results ( more than 90% recognition

rate) were obtained by using features extracted from a

combination of projection profiles in several directions and

cress-sections in vertical and horizontal directions.

Aaaptive learning method improved the performance of

recognition by about 20%.

chapter 2 provices the general background in various

techniques of pattern recognition and chapter 3 describes in

detail the relevant theories and algorithms used in this

study. Chapter 4 describes the implementation methods and

test procedures. Chapter 5 contains a discussion of the

results. Conclusions and suggestions for further research

are discussed in chapter 6.

c n a p r _

T F t 0 Z T I C A L AND CC'nCEPTUal a A C K c R C U N D

The techniques use a to solve pattern recognition

problems can be grouped into three general approaches;

namely, template matching, the statistical (or cecision

theoritic) a p p roach and the syntactic (or structural)

approach. Fi lure 2.1 shows th? ty^icjl statistical pattern

recognition system which consists of a feature extractor and

pattern classifier u s i n ithe- feature measurements from the

input bittern. The syntactic pattern recognition system, as

shown in Figure 2.2, consists of pattern preprocessing,

primitive selection and syntax ar a lysis. Tnjs chapter

prolines an overview of these it e t h o a s in ,_ a t t a r n

recognition.

Tne first step in any pattern recognition problem is to

select discriminatory features representing the pattern and

to extract (measure) these features. For example, the most

important features of nan awritten characters are the

direction of the strokes, the arrangement cf strokes anc the

interrelation oetween the strokes. Tnpse features, in

general" ifi a y not Dc e a s i I y rre.isuracle. Usually, j>-

i r a r y

i.-na^e of the pattern can oe easily obtained and then is

preprocesseo to extract significant features.

OLLX

UJ

a.Q_

i

z\- Pi

z>a.

2__

<

Q

__

0

9

_M

TELUGU Ct-APACT-.' ~.cCC _M T I Li P a q e 7

The features can di classifiac into tnree categories:

(1) physical features, e.g., color (2) structural features,

e.g., ahape, structure and other geometrical properties and

(3) mathenat i ca I features, (e.c., statistical means,

correlation coefficients, eigen values and eigen vectors of

covariance matrices, ana otner invariant properties). It is

generally difficult for a machine tc imitate the sensory

capability of physical features. Syntactic or structural

pattern recognition methods hs^e been developed to analyze

structural features. The oecision theoritic or

cla^sif icatory approach is generally used for p Hterrs where

features are extracted using mathematical methods.

2.1 Template Matching

An irtutive approach to pattern recognition is

"template-matching"[Ros 76], [Hen b4J. In this case, a set

of templates or prototypes, one for each pattern class, is

selected. An unknown input pattern is compared with the

template of each classes and the classification is baser on

preselected matching or similarity criteria. In otner

words, if the input pattern matches the template of the ith

pattern class better than it match as any ether templates,

then the input is classified as a menoer of the ith pattern

class. This apnroach has been_ s e 1 for s:me existinj

orinteu-character recognizers and bark-check readers.

TELUGU CHA^ACTEi- KE'L c-M T I ' i-a ; e t.

Tne following are standard measures of the decree of

match between a picture f(x,y> ana a template T whose gray

I evel at ( x,y ) is t ( x ,y ) :

max of J f (x,y)-t( x,y ) i , x ana y in T (2.1)

ft J f ( x , y ) -

t ( x , y ) J ef.:act

-f<x,y)-t(x,y) ]h..dl<

#

(2.2)

( _. . 3 )

These measures are all zero for a perfect match, and

hr\\ e high values for poor matches. They differ, however, in

the types of errors to which they are sensitive. For

example, if f(x,y)=t(x,y) exceot at few points, where !f(x,Y

- t { x , y ) ; is larje, measure 2.1 will show a large mismatch,

whereas measures 2.2 ana 2.3 will show negligible

mismatches. On the other hand, if ! f ( x , y ) -t ( x , y ) ! is

everywhere non-zero out small, e.g., f(x,y) = t(x,y) + e,

then measure 2.1 yields a small mismatch, e, wnile measures

I. I and 2.2 yield large mismatches e!T! and e2!T!

respectively where i T i is the area of the template T.

Another form of measure 2.4 is the correlation

coe f f i c i e "' !

HT -> fcC^d*.-^

/iST&9)^ ^Ttc.^>H3C*-A)

The value of this function is always between C and 1, with

value 1 being achieved only if f(x,y)-ct(x,>) where c is a

positive constant.

L o G U C H - R A <_ T i-

-'.ELL G N I T I 0 i h a c a v

The disadvantages of the template matching approach

include the difficulty in selecting a good template for each

pattern class and in defining a proper matching criterion.

The difficulty is especially remarkatle wher large

variations ana distortions are expected in all the patterns

belonging to one class twid 7 4 J . The use of flexible

template matching or"rubber-mask"

technique has been

proposed Xj y [wid 74J and [Che 7 5], in an attempt to

accommodate hr,u variations in oattern samples. Also, the

process of matching a template against a picture in all

possible positions is computationally costly. Some methods

of reducing tne cost of terns! at- tr.itchinj Can oe found in

Zar 72 1 a no [Nag 72 1.

1.2 Statistical Methods In Pattern Recognition

In this approach, a set of characteristic measurements

are extracted from tne original pattern anc classification

is dene by partitioning the measurement space into regions

of pattern classes, using dec isior functions derived from

pattern samples. Clustering transformations can be used on

the measurement space in order to cluster the Lcints

representing the same pattern class. Such a transfer nation

will maximize the mean-square cistance Let*, a en pattern

feints that oeloni to two different classes and minimize the

mean- square aistance oetween pattern points of th*: same

class.

FELUuU cHArACTEP RECCGMTIC; P a : e 10

Alternatively, properties of patterns like moments,

projections and cross-sections, can be used as

characteristic features, ano suitable similarity functions

can be chosen for classification.

2.2.1 Feature Extraction -

The basic idea of feature extraction is to obtain

unique information representing the pattern from the gray

level in the pattern at different points.

2.2.1.1 Methoc Of "cments -

In the method of moments used Dy Dudani, et al [Dud 77]

for recognizing aircraft, the picture under study is divided

into cells, and the second moment is computec for each cell.

The second moment (variance) is defined as the sum of the

products of vertical and horizontal positions of gray levels

in a given cell. A pattern then can be represento by

monents. Moments will be the same for identical patterns

anc different for dissimilar patterns. By observing the

values of the moments, it can be determined whether the gray

values are spread out horizontally cr vertically, however

this method is inadequate for describing structural

information of complex patterns [ L _ c ? 3 1 .

T_Lu<ZcH.;AcT_i- Z c 0 ., N i T I 0., Pc.;? 11

2.2.1.2 Projection Profiles -

A gooc ide3 of how the gray levels in a given regior of

a pattern are distributee can oe obtained ty examirinj the

projections of the catterns in various directions. The

projections in x ana y directions are given ty:

j=i

A set of projections in a sufficient numoer of directions

contains enough information to reconstruct the Picture. For

objects having higher gray levels trap their oackgrounds,

;j e _ k _s in projections can indicate locations of major parts

of objects. This is especially true when t n _ image is

normalized to contain only G's and l's. This method, then,

shoulo oe well suited for character recognition. Song [Son

35] and Li ILi S3], used Fourier descriptors of orcjection

profiles in horizontal and vertical eirectiors as features

for Chinese character recognition. A very high success rate

w3s obtained due to the fact that the characters were

corrposad mainly of vertical and horizontal strokes, and the

projections represented these components very wall. In

adcition, the positional invariance of the characters was

achieved by deriving the i-ourier cescriotors, and stroKe

width correction was applied usinj the convolution technique

of the Fourier transform.

In this thesis the projection profiles of lelugu

characters in four directions namely, horizontal, vertical,

-45 a ik< +4s degree, were used as features. A combination of

the four sets cf features yielded unique characteristics.

TELuGU CrAKACTcr-- E C C

Gi"I T I L \ a h a 12

Details of this method are described in chapter 3.

2.2.1.3 Cross-sections -

More detailed information about the arrangement of gray

levels in a region can oe obtained cy examining the

cross-sections in various directions (e.g., the

cross-section in the x-direction is the function f(x,y') for

a particular value of y'). For a binary image matrix with

only Zs and 1's , the cross-sections will si ruly oe. the

numoer of distinct sequences of runs of l's. Figure 2.4

shows an example of cross-section values in horizontal

cirection for numerals U through 9 . Figure <_.:> illustrates

the cross-section values of a Tamil character. Peaks in the

cross-sections correspond to object parts. Comoarisicn of

cross-sections of different pictures can give useful

information aoout object shape in terms of how the peaks

shift, expand, shrink, merge and split.

uutta [Out 7 4] used the change of runs of l's in

horizontal and vertical directions to derive 'events', 'half

events'

etc. in the recognition ot handwritten characters.

Siroir oney [Sir 7 b] used the frtouecy of runs of l's in

columns and rows for recognitior of printed Tamil

characters. The -fcthoc is lescr ibf c celo*:

The pattern matrix is examined column oy column and the

nuiroer of runs of l's is noted for each column. This gives

a sti iiv.i ot numbers. This string is condense.; by celeting

00001111111111111111100

00001111111111111111100

00001111111111111111100

00001111111111111111100

00001111000111100000000

00001111000111100000000

00001111000111100000000

00001111000111111111000000011110001111111111000000111100011111111111000001111000111111111111

00001111000111110011111

00001111000111100001111

0000111100011110000111100001111000111100001111

00001111000111100001111

0000111100011110001111100000000000000000011111

00000000000000000111111

0000111111111111111111000111111111111111111100

001111111111111111110000111111000000000000000001111000000000000000000

111110000000000000000001111000000000000000000011110000000000000000000

11110000000000000000000<X$*22

lo- r__ 11112M222J33333311111111111

Co_-B-- rs. rut 12J1

yaboUe roa ru* -1-8011

C-la-B ru_ 11112222222-22233332211

Coaaana-. oely-a m 12321

jnbolK aolu-B. ra_ -11213*2-1

Fl<i .3CROSS.- BCT\0WS

TO* ,r"

T t L u G u C p A .-. A C T c - Z C C o N I T I On p -a a e 13

the consecutive occurences of the same numbers. This

condensing procedure sncrtens the string considerably and is

equivalent to a sort of thinning procedure. Similarly, the

matrix is examined row by ro and another ccndensed string

is fomed. The condensed strings will oe the same for

identical patterns.

The same method was enhancec to incorporate the

relative lengths of the pattern segments. As in case of

earlier method, the number of l's is notea. In this string,

any one numeral may occur in consecutive positions. Short,

Tecium and loiv^ runs are formed depending cn the number of

consecutive positions of the same numeral. For example, if

the ienath of the pattern matrix is, say, 12, then 1, 11,

111, will form a short run; 1111, 11111, 111111 will form a

medium run and the rest will form a I o n j run. This was

called the symbolic run method and gave unique

representations for Tamil characters [Sir 78].

The symoolic run method, however, is very sensitive to

individual differences of handwritten characters since small

variations in writing can cause large variation s in

cross-sections. however, this method seems to wcrk very

well for patterns with horizontal anc vertical strokes and

ith certain connectivity oroy.rties. For this thesis, the

condenseo run method was examined in detail fcr the

recognition of telugu cnracters, and the details are

described in Chauter 3.

TELUGC cHAt-AcTF- Zt cGrlTIC 1^

2.2.2 Classification Methods -

Mathematically, thp problem of pattern classification

can be formulatea in terns of"discriminant"

or"decision"

functions. Let HI, * 2 ... , w rr be designated as'm'

pattern classes to be recognized. The pattern space, then,

can be considered as consisting of'm'

regions, each of

which encloses the pattern points in a class. The

recognition problem can now be viewed as that of generating

the decision functions, d 1 ( X ) , d2(X), ... ,dm(X>. These

furctions are scalar and single-valued functions of pattern

X . If d i ( x ) > d j ( X ) for some i anc for some j = 1,2, ...m,

iOj, the pattern x belongs to pattern class *i. In other

words, if the ith decision function, di(X), has the largest

value for a pattern X, then X belongs in wi. Such an

automatic classification scheme using a decision -making

process is illustrated conceptually in Figure 2.5.

[ne decision functions can be generated in many ways,

dependiny on a priori knowledge about the patterns to be

recognized. aayers classification rule can be used if the

probability density aistricution functions are Krown [Fu

76b].

Some of th? more commonly useo decision functiors are:

( l) . linear discriminant function,

CECI&ldN

Function

az*;>

)

XSfcCVSlOM

FDNcTlOw

sewOrator

DtOSIO/U

FCJfoeTlO/O

(SiEWEt-iTrOR

^2oo

1

ppoc^zs

X e

11

PArrepw |

fl duCx.

dftooSAMPLES

Dtc\r\oN

*UN)crpou

CLMtl^ilOR

r

11

1

1

1

duwDtnsiow

Fui\)cnow

6EWERATDR.

to,

fc 2-S ,L0C1< t^f\G.pAlA OF- A PA~n>fO CLf\SSllPiE|$.

TELcc-U CrAxACTtv . z LZ G \ i T I u N 15

(2). minimum distance classifier,

(3). piecewise linear ciscrininant functions

(Nearest neighbor)

(4). polynomial discriminant functions.

Another method of pattern classification is to use the

concept cf distance functions [Tou 74]. One way cf

estaalishing d measure of similarity oetween pattern

vectors, is by determining tneir proximity. For example, as

shewn in Figure 2.6, it is intuitively obvious that X

belongs to pattern class Z, solely on the basis that it is

closer to the patterns of this class. However, it might be

difficult to classify Y into either pattern class based on a

measure of the proximity of this pattern to a class. The

method of pattern classification by distance functiors works

Getter when the oattern classes tend to have clustering

properties. Some of the distance functions used are:

minimum distance method,

maxiuin distance method,

K-means algorithm,

Isodata algorithm etc.

A similarity measure or dissimilarity measure gives a

numerical value to the notion of closeness or cist a nee

oetween two objects. The choice of a similarity function

depends on the ease of comouting and its capacity for

discrimination of various objects in the measurement space.

TELUGU CHARACTER RECOGNITION__,

Page 16

Minkowsky metric:

^^-[Jle-Tirj

Cainbe r ra metric:

IR-T.l

d2 CF,T)a- 1 R -vTj |

Chebychev metric:

c*3 CFjT)= MAX | Fj -Tj I

j

Quaarat i c metric:

uj)h.re <5 is o inixn *>os'-+ive. dU-j-LrnL-ti. -Trrcsityv-A

Mahalanobis metric:

a^CF,T)= CcUtwf cf-ti w_1c:f-t)

Cor relation

"Cityblock"

metric:

4O/0 * 2L WjlFj-T,-]

cohere UJ u> cv we^aht vector.

"Ch i-squa re"

metric:

^"--HlS-^fj

where F. ? - 2 F; c\md R cz ^ R

m y.

J-

2, Ft. c1^ F - ^ F"tL-.I -}=( J

F ouv/vJ. I estno The <f>b)ec+c .

X

B

^

PATTERNS cL/,S^F\Ar3LE 137 PRDxi MITY catMcC^rpT

wa

^2^-6 ^VTftRNS NJT CLASSIFIABLE ftf p^lMTy (.A><-FT

T -

LUG UC."

^ r. A c T 'a;. GM T I c N P a g e 17

2.2.3 Learning-

With the linear classification functions described in

section 2.2.2 , perfect recognition is possible with correct

values for the coefficients or weights. However, in

practice, these values for the weights are usually not

available. under such circumstances, the pattern classifier

car have the capability of estimating the best values for

the weights fro>T> the inout patterns. The basic idea is

tiioti by observing patterns with known classifications, the

classifier can automatically adjust the weights to achieve

correct recognition. The performance of the classifier is

suppose; to improve is >nore and more sample patterns are

applied. This process is called learning or training.

For each characteristic of the pattern, in initial

ei^nt is assiqneo. Tnis weight is increased if its

characteristic lean to correct recogntion and decreased

otherwise. The weight adjustment can be done by a fixed

increment or by an appropriate fractional ccrrelation rule

IFu 7fcbJ .

In the early pattern recognition system used oy Uhr

[Uhr b 3 ] , the recogntion proceaure utilized weights or

amplifiers at various levels. The recognition orocecure

involved tahiru tne difference oeteen each characteristic

cf the input pattern and the standard dicticnary pattern.

These differences were then weighted by the corresponding

pattern jtlI ilitTSi and then * e i o n t e d a j, a i n _ y oener j|

T E L u G u C H A h A C T b i- Z C 0...M T I L Fa je 13

amplifiers representing the q v er age of oattern amplifiers

across all patterns, * h i c n finally oroauced a weighted

average difference between the input and the dictionary

patterns. This average differnce was multiplied by a final

average difference amplifer to octain a difference score.

The pattern * i t h the lowest difference score is selectee as

the correct standard pattern.

After a pattern was recognizee, the pattern amplifiers

in those patterns that nad difference scores less than or

only slightly aoove the difference score f cr the correct

pattern, are modified. Tue correct pattern * -j s compared

with each of the si Hilar patterns in turn. Each

characteristic was examined individually, and a

determination made as to whether the correct pattern would

have been chosen if the choice had oeen made on the basis of

this characteristic alone. If this cne cnaracteristic would

have identified the correct pattern, then the correspencing

amplifier was turned up by one. If it would have identified

the wrong pattern then the amplifier was turned down by cne.

If a pattern had a higher difference score than the correct

pattern, then the amplifiers were aojusted only in the

incorrect pattern. Otherwise, amplifiers ere adjusted in

both patterns.

A simplified method of adjusting the weight factors was

used in this thesis ano is describee in Chapter 3 in detail.

Tt-LUCU ZhkAcZ .C'jMik 'age i )

2.3 Syntactic r'.etnods In Pattern kecognitior

In some pattern recognition proolems, the structural

information that describes each oattern is important. In

the case of picture recognition or scene analysis, the

patterns being classified are usually quite complex and

require a large number of features, hence tne number of

descriptors will be large. It becomes impossible to regard

each descriptor as defining a class. Consequently, the

requirements of recognition can be satisfieo only by a

description for each pattern rather than by the simple task

of classification. Some Isnau^ge scripts like aevanagari,

Telugu and Chinese, with thousands o r character., fall into

this category.

Une met hoc of representing the heirarchical (tree- like)

structural of each pattern is to describe the pattern in

terms of simpler subpatterns, with each simpler subpattern

a g 3 i n described in terms of sven simpler subpatterns, etc.

The simplest subpatterns selected, called "pattern

primitives,"

should be much easier to recognize tnan the

patterns themselves. The"language"

providing the

structural description of patterns in terms of a set of

pattern primitives and their composition opearations is

^uirf.tiiivs called a "pattern descriptionlanguage."

T h a rules

governing the corrncsition of primitives intc patterns are

usually specifiea by the % r a m rr a r"

cf the pattern

description language. After each primitive within the

pattern is identified, the recognition process is

T _ L U G U c H A i- A C T _'.

'

l c c GM T I C :. p o e 2 0

accomplished by performing a syntjx analysis or parsirg of

the"sentence"

describing the given pattern to determine

whether or not it is syntactically (or grarratically) correct

with respect to the specified grammar. In addition-, the

syntax analysis also produces a structural descripticn of

the sentence representing the given pattern (usually in the

form of a tree structure).

J i~\e various relations or comoosition operations defined

j;ron; subpatterns usually can ce expressed in terms of

logical ano/or mathematical operations. For example, if we

cnoose"concatenation"

as the only relation (composition

o aeration) usee in describing > a 1 1 e r n \. , then for the pattern

primitives shown in Figure 2.7, the rectangle shewn in

Figure 2.8 would oe represented by the string 'aaactccccd'.

vore exolicitly, if * e use"?"

for the he.a u-to-tail

concatenation operation, the rectangle in Figure 2.6 would

be represented by the string 'a+a+a+b+b+c+c*c+d+d' and its

corresponding tree-like structure would be as shown in

Figure 2.9. Similarly, a slightly more complex example is

given in Figure 2.10 using the pattern primitives jiven in

Figure 2.7.

An alternative representation of _ pattern's structural

information is j "relational < r a o n". In using p. relational

graph for pattern description, one c^n ircluca any relation

tnat can oe conveniently determined from the patterr. Note

that (1) concatenation is t n e only natural operation for

one-dimensional languages, and (2) a graph can contain

BjJ2-7- P.VTnlHV^

_x<** ex.

TZ S-fl. P> fe-thsw^gU. wfrfc

u

toiiW Kyf> W, Pi 2-7-

Rectb_n<\b

<=*-i-d.-t<a-i-b + b-V c_a-

C + -_ +_^ -t <A.

a-V b

Fu j_r*-& Tjlbc <%i>vaj->tuA_: -fee

bj>tA.fr^4>V.''/> W tr -2.. "7

-\-*l ,

C ,C

k fc

* V*

c c

9

c -+__ + dl + a -t-a-t- b Mb*-c + c_

^Q 2-.10 fcvX-u- 9 a**il y-W

TELUGU CHAKACTc< ^KCCGMTIuN Pa ,e 21

closed loops whereas a tree cannot, tnerefore, graphs can be

used to express richer cescriptiors than trees. However,

the use of tree structures does provide a direct way to

adapt the techniques of formal language theory to the

problem of compactly representing and analyzing patterns

containing a significant structural content (Fu 76], whereas

graphs do not provide a direct way of analyzing the

structural content of a pattern.

A labeled graph aporoach was used to represent the

structural anc connectivity information of strokes in the

Tamil script for handprinted character recognition by [Chi

o 0 ] . In tnis approach, the characters were assumed to be

domposed of line-like elements, called primitives (Figure

2.11a), satisfying certain relational constraints. Cirected

labeled graphs are used to describe the structural

composition of characters in terms of their primitives and

the relational constraints satisfied by them. For example,

the nodes of the grapn in Figure 2.11c correspond to the

junction names of the letter'

8>'

' n Figure 2.11b. A

junction can either be an end point of a primitive or a

point where two or more primitives Teet. The label and

direction of an edge represent the primitive joining the

junctions named by the end nodes of the edge. For anlysis,

tne jraph of a letter is represented by a connectivity

matrix. The recognition algorithm uses a topological

matching procedure to compute and maximise the correlation

coefficients. basic symbols are first recognizee to

Pf_MM\nv--_

1

b

- X A

<

f

>h

SMI*.

"JUNCAOM

12.3

Gio

2.|lb -rHACACTtl? Cwa)

fch'i 8c0

*) v || io

2. lie

dmtcrtro lp,cz UP ^ te*tn> or

TcLUGU cHAkACft-J ->-c:Ci. CM TI _: P a e 2 2

identify derived symbols. An excellent recognition rate was

obtained by using method.

2.3.1 Linguistic Pattern Weccgnitior Syst ems

A linguistic pattern recognition system general

consists of three major phases: preprocessing, pattern

description or reoresentation, and syntax analysis, as shown

in Figure 2.12. The functions of creorocessinc include:

(1) pattern encoding and approximation, and (2) filtering,

restoration and enhancement.

In the preprocessing unase jn input pattern is first

coceo into or approximated by some convenient form for

further processing. For example, a b 1 ack-an c-wh i t e picture

can be coded in terms of a grid or matrix of G s and l's,

diic a waveform can be approximated oy its time samples or a

truncated Fourier series expansion. Data compressior is

often appliea at this stage to make the processing in the

later stages of the system more efficient. Then, techniques

of filtering, restoration and/or enhancement are used to

clean up noise, to restore tne degradation, ano/cr to

improve the quality of the coded or approximated patterns.

The output of the preprocessor, then would ce patterns -ith

reasonably "good duality". E:<ch pre processed pattern is

men represented by a sentence-line structure, e.g., a

string or a graph.

u

2O

E

0

sfl-

X

^

ft

9CQ

k

(N

_?

u.

TcLUGJ ChAr. ACTZ Z\I riuii P i ,1 e 2 3

The pattern representation process consists of: (1)

pattern segmentation, ano (2) primitive (feature)

extraction. Each pre processed o a 1 1 J r n is segmented into

subpatterns and pattern primitives based on prespecified

syntactic or comoosition operations and, in turn, each

sutpattern is identified with a : i v e n set of pattern

primitives. taoh pattern is then representee oy a set of

primitives with specified syntactic operations. The

decision on whether or not the representation (pattern) is

syntactically correct (ie. eel onus to the class of patterns

described by the given syntax or grammar) will be performed

oy the syntax analyzer or"oarser"

. The syntax analyzer

usually can produce a complete syntactic description of the

pattern in. terms of a parse-tree, provided the pattern is

syntactically correct. Utherwise, the pattern is eitner

rejected or analyzed on the oasis of other given grammars,

which presumably describe other possiole classes of patterns

unoer consideration.

The simplest form of recognition is "template

matching". The string of primitives representing the input

pattern is matched against strings of primitives

representing each reference pattern, and the input is

classified in the same class as the reference pattern which

is the"best"

match. '*ore complex recognition rrethods

involve a complete parsing of the input string exploring the

corrplete heirarchical structural description of the pattern.

The selection of appropriate approach for recognition

Tt-LUGU cHAi-ACTb"-"cCLuMTZiN ge 2 4

usually depends on the proolem requirement.

Obtaining a grammar that describes the structural

information about the class cf patterns under study requires

a grammatical inference machine which infers a grammar from

a given set of training patterns in language-like

representations. This is analogous to the"learning"

orocess in statistical pattern recogniticn systems. The

structrual descriotion of the class cf patterns under study

is learned from the actual samole patterns fro'i that class.

The learned description, in the form of a grammar is then

used tor pattern descriotion ana syntax analysis.

Practical applications of linguistic pattern

recognition include the recognition of English, Chinese ICha

73], and Oevanagari characters ti>ir 79], spoken digits,

mathematical expressions And 66J, the classification of

bubble-chairber and spark-chamber photographs CSna t>ti] [Bha

72], chromosomes and finger print i .rages [Co a 75], and the

identification of machine parts [V_m 73J.

An excellent example of the use of the linguistic

method is PLANG, a language descricing e v a n a y a r i script

(Sin 831. This script consisteo of characters with

vertical, horizontal anc cursive strokes. The patterns are

aescrioed in a nattern description language for recognition

ano analysis.

T_LUGc CHA-aCTcK -. LCLGMTIIJ., P d c e 2 5

A sentence in PLANG aescriceo a two-dimensional pattern

in terms of primitives and composed macros specifying their

relationships with respect to various regions of the picture

frame or in terms of subpatterns operated by a frame

function(s). A partitioning fucntion was assurrec which

partitioned the picture intc nine equal regions as shown in

Figure 2.13. Every partitioned region could oe oartitioned

further to get finer'regions'

oy recursively applying the

partition function on a partitioned region. The urion of

regions could be defined by using the U operator. A

sentence in PLANG was representee as the outcome of a

frame-function'

operating 00 on a list of'frames'

or

another PlaNG sentence. Two or more picture frames could be

combined in a desired manner to mane 1 larger picture frame,

or an original picture frame could be transfer ne a to a new

picture frame in a desired manner utilizing a set of

frame-functions. Two frame functions narrec'superimpose'

(denoted by' * '

I and'append'

(denoted oy *.*) were used for

Oevanagari script. HP.Z (horizontal line), VERT (vertical

line), LTD (left-going diagonal), w TC ( r i gh t-co i n 3 diagonal),

CURVE (arbitrary curve passing through specified re iiors in

that order), and uUT (a dot) are selected as the primitives,

and were illustrated in Figure 2.13. Compostion of the

Oevanagari character kRAIn and itj PLAr.G description are

.,iven in Figure 2.14.

LT

1

1 MT! RT

LM 1 MM ; RM

fLB i

i

MB ! RS

Fig. I. Partitioning function.

(MM MBA)

Fig. 2. Operator V and repanitioning.

Pi

a.

VERT

Ml

LTO

DOT

HRZ RIO

I . initiol region

2 _ terminal region

Fig. 3. Primitives.

2.-Y& Bt-tLP PR\K\TiUES IN PATTET5.IU &ESCR\PT1CI0 t/VKi(=cOA<5.E P_>\WG-

L*-JU) o ())^*J

.

LLJJ l- ) --Bl I

OK(ltT) 0 (ii)-).

S

uuti)

(if )"

8ayrfp

tint

isir

n

13 -S ItiperlapM*

DOT ((

DID

lupartapM*

S-((-t) 0 (UM)

f-

i

Supeftapoac.

C0RY1

(M1(MM B))(_ I) 4a

"'SJ tuperlmpoe*

LTB(m B)(U I)10

J!

ch

~-_ ~0.0.4J) /

rS-i12

U-X

I IvperlapM* 00m(m i)((_T LH) U (If U)B)((LN M) 0

(MR U)B)((M U) 0 (NT ->)) (Ml I)

Composition of Dcvanagari composite character KRAIN.

DEVKRAIN

Describe KRAINDescribe KRAI

Compose

Compose

Compose

Compose

Compose

Compose

n

Begin

(.((KRAI)(DOT((MT) U (RT)ll))))((.(0 0 0 0) (0 .45 0 ,45))((KRA ((LM) U (RB)0))(AI ((LT) U (RT)Q))))((KRA(R))(.((KA(R))(LTD(MM R) (LB R)))))((KA (R)) (.((HRZ(LT R)(RT R)) (VERT(MT R)(MB RECURVE (MM R)(RM R) (RB R)) (CURVE((MT MM) U (LB MM)R) ((MT LM) U (RT LM)R) ((LM LM) U (MM LB)R) (MM LB)U (MT MB)R) ((MT MM) U (LB MM)R)))))((AI(R))(.((CRL (LT R))(RTD (LT R)((MB) U (RB)R))(CRL (MT R))(RTD (MT R) (RB R)))))((CRL (R))(.((CE((MT) U (LB) R))(DE(MT) U (RB)R)))))((CE(R))(CURVE (RT R) ((MT) U (RB LT)R) ((LM) U (MM)R)((MB) U (RT LB)R)(RB R)))((DE(R))(CURVE (LT R)((MT) U (LB RT) R) ((RM) U (MM) R)((MB) U (LT RB)R)(LB R)))(coordinate- of lower left corner or picture-frame height width)

End

D^\/ov.vv.Cva'vrv_ C^o^^\-C.CcrL

TcLUGU O'AkACTE r. J ti C i, i\ I T I C r Pace 2b

A PLANG description for all synools of the script were

stored. A local feature extraction operation was performed

in which every point of tne pattern was assigned a label

depending upon the local property it exhibited with respect

to its neighooring points. The local properties included

wnether the point in question forrrea the part of a vertical

line, horizontal line, rignt-going diagonal, left-gcing

diagonal or do not care. These labels were used by a goal

oriented top-down parser for primitive recognition. In

Figure 2.1b, a hierarchic representation of the PLANG

description for the symbol KA is s n o w n . The parser looks

for tne fr^me function and Iocks for the appropriate

primitives in the pattern. If it is successful, the parser

moves on to other suotrees towards the right. If, at any

stage, the goal is unsuccessful, a penalty is enterec and if

the penalty exceeds a preset threshold value, the carser

picks up the description of the next probable syrrbol.

Contextual information about the script helpea acnieve

a recognition rate of more than SO percent.

2.3.2 Selection Gf Pattern Primitives -

The determination of primitives with which tne oat tern

of interest may be described is largely irfluencea cy the

nature of the o a t a J there is no general solution for the

primitive selection problem tFu 76]. The primitives should

serve as oasic pattern elements to (.rovide a corrpact out

adequate oescription or tne data in terms cf the specified

(>)

1 I_l

m> i i i io

J_

T

T

J_

I

( I IB

(b)

Fig. 6. (a) Thinned patlera. (b) Labeled pattern.

KAIaj)

< n n vERTn n "nnIT A RT A MT A MB A MMAR8A

Fig. 8. Parsing tree for symbol "KA".

CRL

MM U LM A

2->D HtlRARCtflC ReFREcer^TA-noNi pOR. ^tASOL

y

CSin S"5l

T E L c G c C H A r; A C T c LCGMTl c N Page 27

structural relations (e.g., tne concatenation relation).

The primitives should be easily extracted cr recognized by

existing non-1 inguistic methoos, since they are considered

to be sirrple ana compact patterns and their structural

information is not important. For example, for speech

patterns, pnonemes are naturally considered as a"good"

set

of primitives with tne concatenation relation ICho ti] .

Similarly, strokes have been suggested as primitives in

describing handwriting [fcde bl]. however, for general

pictorial pat terns, there is no such universal picture

element.

Freeman's chain code [Fre blJiFre b 2 ] is j common set

of primitives used to descrioe the boundaries or skeletons

in a pattern. In this scheme, a rectangular grid is

overlaid on a two-dimensional patter r and the straight line

segments are used to connect the grid points falling closest

to the pattrn. Each line segment is assignee an octal digit

according to its slope. Patterns are thus represented by

chains or strings of octal digits. Figure 2.16 illustrates

the primitives and the codec string descricing a curve.

Si rr pie manipulations like rotation, expjnsion, measurerrent

of curve length, and determination of pattern

self-intersections can be easily carried out. < r o k e and

Wiley IKno o 7 J ind Feder IPed t.Z used this 'net hod tor

uescrioing hand-printed characters.

T t L U G U ChiKAiT^ rt-Ll.&MIl Pa:e 2 8

A set of primitives encoding geometric patterns in

terms of regions has been prooosea by IPav 6 5 1. In this

case, the basic orimitives are halfp lanes in the pattern

space. It can be shown that any figure (or arbitrary

polygon) may oe expressed as the union of a finite number of

convex polygons. Each convex polygon can, in turn, be

representea as tne intersection of a finite nunber of

halfplanes. By defining a suitable ordering (a sequence) of

tne convex polygons composing the arcitrary polygon, it is

possible to o e t ermine a unique minimal set of maximal

polygons, called a crimary subset, the union of which is the

i_h I v e n oolygon. As a linguistic analogy, i fijure can be

thougnt of as a 'sentence', tne convex polygons composing it

as'w o i d s ,

'and the halfplanes ^ s 'letters'. A more general

selection procedure cf pattern primitives based on regions

was proposea oy s<osenfield and Strong [Ros 7 1].

Another form of representing polygonal figures is the

use of primary graphs IPavl 7 2 1 I P a v 2 721. The primary graph

of a polygon A is one whose nodes correspond to the nuclei

and primary subsets of A, and whose branches connect each

nucleus to all the primary suosets containing it. An

example is given in Fi jure 2.17. Primary subsets anc ruclei

of polygons approximating the f iiures are shewn in Figure

d. i / i (shaded areas are nuclei). Primary ; r a o h s for the

corresponding polygons in Figure 2.17a are given in Figure

2.17b. This form of representation provides information

about the topology of the picture- Also, patterns

M L

Primary Subset

or Nucleus

Label

ABLM a

ADEP b

CDH1 e

OFCN _

ABQP 1

CDER 2

OTKN 3

SFGJ 4

%j_Zf

_-_y

a b

FiejJ_?i |*j_) Polygonal figures and b) corresponding primary graphs L. Fo.\j; _L T2.J

7Zt

5 fv.*

Coded siring of llie curve 7600212212

b

fig. 4.10a and b. Freeman's chain code

_? |jj,\-zrcsssLSrJ\ c*/vo->

;h=xir> code Cfvsui. 6-2.J

TELUGU CHAKACTt^ ZZGMTIZ P a j e 2 9

represented by graphs can oe forrrally aescribed by graph

grammars which can be analysed.

TELUGUcHAftAcTEx'

KcCCuMTIl.s Page 30

2.3.3 Pattern Grammar -

Once the pattern primitives are selectee, the next step

is to construct a grammar that will generate a language to

describe the patterns under study. we will consicer two

classes of languages, namely, finite-state versus

context-free and context-sensitive. Finite state auotomata,

which are easy to implement recognize or accept finite-state

languages. However the descriptive power of finite-state

languages is weaker than that of context-free and

context-sensitive languages which require non-de t er rr i n i s t i c

parsing procedures. A precedence language may be used for

pattern oescription in order to obtain efficient analysis.

Cn the otner hand, a context-free programmed grairmar

generating a context-sensitive language may be selectee in

order to describe the patterns effectively.

As an example consider the language

which could be interpreted as the language describing

squares of side length n =1,2,...

C c

nOl

L is a context language anc can be generated by a

context-sensitive grammar or a context-free programmed

grammar as shown in Figure 2. IB.

The language

L={(fbnd,d"\n^{}

could be interpreted as the language describing squares of side length n= 1, 2,.

c

d{ \b

a

o

L is known as a context-sensitive language, and can be generated in the followingtwo ways.

1) A context-sensitive grammar

G.=(KN,V'T,P,S)

where

VN= {S,A,B,C,D,E,F,G)

Vj={a,b,c,d)

P: S-*aAb dG-*Gd

A-*aAC aG-mbcD

A-+D bG-rbbcD

Dc-+cD dFB-*dFd

Dd-*dD dFd-*FddDC-*EC cF-*Fc

EC-*Ed bF-*bbc

DB-FB aF-*ab

Ed-*Gd bB-*bcd

cG-*Gc.

FlCiV 2. I'ScxCerv-,h.^-t-

-/SiOv^v'rLv/G. g5ovwv\M.cCTLF<-^~?6j.

where

A context-free programmed grammar

G2=(VN,V1,P,S,J)

VH={S,A,B,C,D)

VT={a,b,c,d)

J-= {1,2, 3, 4, 5, 6, 7}

P: Ubel Core Success

field

Failure

field

1 S-aAB {2.3} WA^aAC {2,3} {B}A-D 14} {fl)C-d {5} {6}D-*bDc {4} {fl)B-d {') (0)

. Dbc {0} (0)

n* 2.* b. ^t-f.p,j^^

a~~~- ct_.^a

TELUGUCi-AnA-Tlr*- k cCC GM T I L .n P^ge 31

Although many classes of patterns intuitively appear to

be context-sensitive, context-sensitive grammars have rarely

been used for pattern description simply because of their

complexity- Context-free languages have been csed to

oescrioe patterns such as English characters, chromosome

images, spar k- chamber pictures ISha btf], chemical

structures,finger-

print patterns ana spoken digits [And

68]. Figure 2. IS a describes a context-free grammar

Gescrioing the cnron, csome images shown in Figure 2 . 1 S o .

In patterns using string grammars, the only relation

oetween subpatterns and/or primitives is concatenation; that

is, eacn sucp ittern cr primitive, can be connected only at

the left or right. Tnis one- dimensional relation is not

very effective in describingtwo-

or three-dimensional

patterns like mathematical expressions, pictograms etc. An

attempt was maae to include more useful relations by [Nar

70]. As an example, T R I ANGLE ( a , b ,c ) means that a ternary

relation TRIANGLE, is satisfied by the line segments a, b

and c, and ABGVE(X,Y) means that X is above Y. Similarly

the mathematical expression

ex. + b

c

can be described by

Ao!JVE<ABUVt:(LEFT(3,LtFT( + ,d) ) , ) ,c)

where lEFT(X,Y) means that X is to the left cf Y. A grammar

clesci iomq tne houses in Figure 2.20a is shown in Figure

2.20t>.

V^= {(submedian chromosome), < telocentric chromosome), (arm pair).

(left part), <right part), <arm>, <side>, <bottom>}

VT= \

and

ft l m }

P: <submedian chromosome) - <arm pair) <arm pair)

<telocentric chromosome) -? (bottom) <arm pair)

<arm pair) -? <side> <arm pair)

<arm pair) -> <arm pair) <side><arm pair) -? <arm> <right part)<arm pair) -? <left part) <arm>(left part) -* (arm) c

(right part) -? c (arm)

(bottom) -? b (bottom)

(bottom) - (bottom) b

(bottom) -? e

(side) - b (side)

(side) -* (side) b

(side) -* b

(side) - d

(arm) - b (arm)

(arm) - (arm) b

(arm) -* a

_L.I7c_q7t_^*AWOJT <dbLoc/\A- bl_v^q

z\^b- Zf^i&J:

cUTov\AOSovv\e vwv>v

T

bobcbabdbabcbabd btbabtba

a b

O IC3 t_ a) Submedian chromosome and b) telocentric chromosome

TELUGU ChUKACr.r c C C GM T I 0\ Page 3_

[Sha 701, by attaching a'head'

(hd) anc tai I ( 1 1 )

to each primitive, has used the four binary operators +,X,-,

ano * for defining binary concatenation relations between

primitives (Figure 2.21). Each pictorial pattern can be

representee by a 'labeled or anch-or i en t edgraph*

or

relational graph, nere nodes represent subpatterns or

primitives and branches denote (binary) relations (Figure

2.22). Fecer [Fed 71] has formalized a'

p I e x'grammar which

generates languages with terminals having an aroitrary

number of attaching points for connecting to other

primitives or sunpatterns. Pfaltz and Rosenfield have

extended the concept of string grammars to jramirars for

labeled grapns called'webs'

[Pfa 711. Each production

aescribes the rewriting of a graph A into another graph B

anc also contains an'embedding*

rule E which specifies the

connection of B to its surrounding graph (host web) when A

is rewritten. Here, the terminals or primitives are

represented as vertices in the graph.

2.3.4 Syntax Analysis As Recognition Procedcre -

After selecting the appropriate pattern primitives ana

their suitable pattern gramirar representing the

concatenation relations of the orimitives, the next step in

syntactic pattern recognition method wouIg be to perform

syntax analysis. Syntax analysis or parsing is necessary if

a complete description of the input pattern is required for

recognition. As mentioned aoove, finite-state automata

hd

a + b b

a

a

a-b

a-b

tt ^- -

7?hd

hd(fl+ 6)= hd(6)

tl(a+b)= tl(fl)

hd(flxb)= hd(b)

tl(flxi>)= tl(b)

hd(a-b)= hd(fl)

tl(fl-t)= tl(fl)

_* u ; ". hdhd(u,b) = hd(u)

tl(a._<) = 11(a)

2-2> Cc^ocxJZ-vvxVic^ ^bcHo^s Uitoj;..^k^^Z>/w [>~~Gl

/T

ZTriangle +

House

^\ /tv

Triangle

b * <

'AA

PDL structural descriptionof/4"

and"House"

q^cZh-S- t_F-cl6_l.

cJ-> -c~s i-6^CN/v/Vd") -CUjtA

The following grammar will generate sentences describing houses

G=(VN,VT,P,S)

where

^N= {(house), (side view), (front view), (roof>, (gable),(wall), (chimney), (windows), (door)}

KT=D.Q.U. EH. A. d.^. -,(.). 1

IO. 0.1.-

S= (house)

P: (door)- g

(window) -03, (windows)-- ((windows), 03)

(chimney) - ?.(chimney)- Q(wall)-?, (wall)- O door>, ? )

(wall)- ((windows), | |)

(gable)-A. <gable>- [ ((chimney), A)(roof)-W (rooO- i ((chimney), y\)

(front view)- { ((gable), (wall))

(side view)- | ((roof), (wall))

(house)- (front view)

(house) i ((house), (side view)).

The notation

(X, Y) means that X is to the right of Y,

O (X, Y) means that X is inside of Y,

O (X, Y) means that X is inside on the bottom of Y,

I (X, Y) means that X rests on top of Y,

('

(X, Y) means that X rests to the right of Y.

'2->'2QcX- ^Tscwvwwtct. cto.-)C-v-bL-v- F-jcnjo^c-) <->v\ 2-20b' LT'^-1- ~7--> J

House Description

_j

a

1 (_-.?)

i(i(iA).o(g.D))

Aa

S3-(l(l(D.^).0(_H,n)).l(A.O(Q.Di))

t-Vjt 2 ZO b Pc*-ttt3VT-i du-oCZi-Hcnri cj) heuy/j L FZ "7<oJ

TcLUvjU UHAi^A.Tb^. i-'i-CCGM TZ i >a _ e 3 3

recognize or accept finite-state languages. If a class of

patterns can be described oy a finite-state language, a

finite-state automaton can then be constructed to recognize

the strings or sentences describing this class of patterns.

when a context-free language is used to describe a

class of patterns, the corresponding recognition device is,

in general, a non-deterministic pushdown automaton. The

output of the syntax analyzer usually incluaes not only the

aecision of accepting the string generated by the given

grammar, tut also the derivation tree of the string, which

in turn, gives the complete structural description of the

pattern.

2 .4 Descriptive Characteristics Of Telugu Characters

The telugu alphabet, with more than 2CU0 characters,

consists of vowels, consonants, and consonant vowel

coiroi nations (C-v letters), and combinations of consonants

ano/or C-V letters (conjunct consonants). There are 16

vowels (Figure 2.23a) and 36 consonants (Figure 2.220). A

vowel following a consonant takes on a different graphic

form called a 'vowelsign*

[caj 771- Vowel signs

corresponding to the vowels in Figure 2.23a are sho*n in

Figure 2.23c. The vowels anc consonants combine to torn 57tj

oifferent C-V letters. The C-v letters are formed oy adcing

vowel signs to consonants at approoriate places. Figure

2.23d snows some C-V letters obtained ty conic in in j the first

consonant in Figure 2.23b with all the vowels. Similarly

<9 65 a -6s 6 &m _>_?

Fm. I. Vowrk

'8~

Fits- 2.. -23 ex Vowjels

OOJSci-J <3 d C- V5&

# <3 # ,3 _5 $ EJg^

Fiu. 2. (jJii.M.iianl.s,

2. ---.3 b. ConsoncxnTS

y-

Fi. 3. Vuvd signs (>liirrel vowel *ici_< m-riir only with the nsnimnls :i, !l, 12, 17, 18, 1!), 'ifi,atid 27;.

F.cx. 2-_23c- ^UOfcl Sipjns COV r&sf*=>Y*_t-v>, -ft) VOiveis U. -fo223ex

o

2- 2-3 -A

<ipdjjbpcaJpc^cV

^ C# &>o ^S

tt, Q-2.5dL^e. lW

of C-V kita-. C^ 17]a

TELUGU CHAKACTF^ PcCCGMTIlm Page 34

the characters shown in Figure 2.2ie are obtained using the

ninth consonant. It is clear from the above description

that vowels, consonants, and C-V letters can be realizec by

superimposing snapes called"build-primitives"

at

appropriate places over certain "basic-letters'*. The basic

letters ana build-primitives are shewn in Figure 2.23f and

2.23g, respectively.

C-V letters and consonants may be combined in many

different ways to produce thousancs of different characters

called 'conjunct-constants'. In general, one C-V letter and

one or nore of the consonants combine to form a conjunct

consonant. Conjunct consonants * i t h more than one consonant

are very rare. If such cases are excluded, a conjunct

consonant will consist of two components. The first

component, called the "main component", is a C-V letter.

The shape and location of the second component depenos on

which consonant it is. For some consonants, special

symbols, called'conjunct-primitives'

(Figure 2.23h), are

substituted to form the second component ( F i g ur e . 2 . 2 3 i ) . In

some cases, the corresponding consonant (without the first

vowel sign) is written oelow tne main component; some

examples are illustrated in Figure 2.23J.

c .t> Previous work In Recognition Uf Some Indian Scripts.

kajasekharan and Oeekshatulu txaj 77] have successfully

usea structural methods for the recognition of Telugu

characters. Considerable worn was oone in recognition of

1 23 A 5 a -7 S s

& s s n o x, eotra1> " '2- "* I* '5 16 I? f$

iq _-o

a 22- 25 24 3,5

F,3' ^ 234 Basic Le-Hers.

*""

**"

=r-o of\

r\ 9

_/*-^-*wJJxu.__

'

(}' 2,2^Q- I,W-lriutiiivw.

^Q- 2-2-3 h ONljllllK|.|lri|||i|iv(K.

4 d con. z> gij &Rg. 223

,^i:.v.hi|l,-s of t.i.i.ii.rt .^Ksonaiils will, ...i.j..h.-l-,,rii.,iliv.S<.

rl <5 60 6 a)n cpj w o n

Fja "_>.2.3jW'''x:""I,'('> "' O'lijo'11'' i'uiisi>ii:imI> in wliicli roiiMjiuiriLs wil limit llic- fii>tl vowel si^n

opinair riflow (' V It-llcr..l^aj 77_] ,

TELUGU CHAKV-TL-r-'^CCuMTU'

Pd-.e 35

other Indian language scripts like Jevanajari, Bengali and

Tamil. Siromoney and Chan dr a sekha r an [Sir 78] used a

condensed run method for recognizing printec Tamil

characters. A labelled graph representation for Tamil

characters was used by Chinnuswamy [Chi 8C1 . Sinha [Sin

79] used syntactic approach and a picture language to

oescribe the build primitives of Oevanagari characters. A

knowledge basec system was developed later on to recognize

words in the Uevanagari script [Sin t 4 ] . A generalized

formal approach for description anc analysis of major Indian

languages was proposed by Datta LL.at 1 5 ] . A combination of

Structural anc statistical methods fcr recognition of Tamil,

P a I a y a I a m ana 0ev3nagari scripts was attempted by

Chandrasekharan [Cha 851. Ray and Chaterjee [Ray o 5 ]

designed a classifier for Bengali characters Daseo on

nearest neighborhood technique. A character recognition

system, based on the contextual information in composing the

Tamil characters, was developed by Chanar-isekharan [Cha 31.

A heirarchical decision tree classification scheme is used

by Marwah et al [Mar 84], based on the fact tnat characters

in Oevanagari script can he constructed using certain basic

primitives. (Note: Bengali, Malayalam, Tamil and Telugu

are regional languages in Inoia and have their own

Distinctive scripts. Oevinajari script is used for

languages like Hindi, vl a r a t h i , Sanskrit etc.)

2 . t> . 1 keccgni tion Uf Telugu Characters Using Structural

Methods -

TELUGU LFiAkACTFi; -^lCCGMTICN 36

After a careful analysis of the Telugu characters, it

can be found that all the possible Telugu characters can be

realized by superimposing certain primitive shapes over 25

basic letters. Based on this observation Rajasekharan et al

LUaj 77] devised a two-stage recognition system. In the

first stage, a liven pattern was examined for tne presence

of primitives. If they were found, they were removed from

the given pattern after their presence was noted. The

primitives were looked for in the character in a cefinite

order since some primitives occluue others. Also, the

search for a particular primitive was lirited to a

particular region of tne pattern. After the removal of the

primitive shapes, the remaining basic letter *as recognized

using anon- the- I ine coding procedure. Overall recognition

cf the given characters was accomplished by using a cecision

tree which made use of the primitives and basic letters thar

were present in the input pattern, and contextual

information for composing the characters.

A block diagram in Figure 2.2<t describes various

subsytems of the Telugu character recognition system. The

digitized images were segmented tc seperate composite

characters into basic character primitives. A histogram of

the number of figure points present in each row of the inout

cittern was usea for this segmentation. In a Dinary pattern

of O's and l's, figure points were represented by l's and

background points by O's.

INPUT

Pmn

\>f^FTAOVA-.

,

DIGITIZER!-~

T7VnOfO

TniMf,ifs/_r

x I* A 1ClA^e, > LPATTEfcjN

.

^

LC 1 ILk

-

i*

Pft',WINE

RtCQ-<i-J12L-rCkKsWTR.

^_.

Fl-V ._H TELU^J CLJ-rVVOj^O-ETg. JSECOGr^lTTOM ^ -.Tfc^ LF^rc\ul

FOP- "TR.Acmn.<-

n1.; nl

Fl- [5,7,6^

XI v ? 3,4,5,6?

T--_iui-i ?._. cxeD

Fop TftACUOGt TH

k'JILD PR\Mm\it V

-> succzm

l^zC2-~zed potrrls

e^coorjre-TiE"D

<fi- 1\\\1 TEiACE fc Jjdes (JOT FlfOC

poia;t in Tm; A-.ioK-^JP

DtCitCTiOiUS SPrZo p\fc& JW

A 7-"IL?pLE

^i^n - TU-KJCTiO/J poiMTS

E.f0cO(JMTt ft LTD

F\& C2..X7 TL msD RU.LD pRWMTnjfe V /V^D >rs T-TUPLgS- C&xj"n"3

TcLUGU CHARACTER ZcCGMTIZ, y a g e 3 7

Noise was removed in two phases. In the first phase,

if the number of figure points in the 8-neighbourhood

(Figure 2.25) of a background point .as greater than 5, the

corresponding point was filled with a figure point. In the

second phase, if the number of figure points in the

8-ne i ghDour hood ot a figure point was less than <i , then the

cor r e spona i ng point was removed. The thinning algorithm

developed by Deutsch [Qeu 72J, was then used to extract the

skeleton of the image.

The sequential template matching procedure makes use of

a directec curve- tracer which traces a thinneo-line pattern

by choosing successive points in cne of the allowed

directions. The set of allowec cirectiors is an ordered

subset of the set of eight principal directions (Figure

2.2t>). Let the eight principal directions be denoted by Nx

= (1,2, ...8}. Let R = LR1,R2, .. .RT>, wnere 1<=T<=8, and Ri

belongs to Nx, K=i<=T. The orcered set, called the

T-tuple, specifies the allowed directions in which the

tracer should move. At any point, the tracer looks in the

allowed directions in the order in which they are listed in

the T-tuple and moves to the first cntraversed figure point

encountered. In this way, successive points a r e chesen to

trace the line pattern using a T-tuple. As soon as the

tracer meets a situation at which there is no untraversed

figure point in the allowed directions, it rrakes use of the

next T-tuple given and traces the pattern. For example, the

T-tuples Fl=07,b> and D1 =L3,5,4} are sufficient to trace

KinUl priuciruil dii-uclioii..

Ro.2.25 Kx:uii|iIim of |ial.U'ii_< I Intl. rim Imi Iniml liftingtin- 7'-hi|ili>x A'l iiml l>\. L.'^'u. 7_0

lELUGU ChA-iACTEK *t.CCGMTIC\ Page 3 3

anc recognize the line patterns in Figure 2.2b. The tracer

starts at x and continues up to y using the T-tuple Fl. At

the point y, the tracer does not fine a figure point in the

allowed directions specified in Fl. Hence the tracer gets

the next T-tuple DI ana traces the pattern. At point z,

tracing is stopped.

This method is not affected by variations in size,

stretching or saueezing of the line patterns in the

horizontal or vertical direction, and is impervious to

rotation to a certain extent. however, this method has to

oe aided by other facts like the relative lengths of line

segments traced by Jifferent r-tucles, the region of

operation of the tracer while using a T-tuple, the relative

positions of end points anc junction points, etc. Alsc, the

ordering of the directions in a T-tuple should be done with

great care. Figure 2.26 shows a build primitive, the

T-tuples used, and a flow diagram to direct the tracer.

For recognition of the basic letters, an

'on-the-cur

ve- t rac ing'

procedure was used [Raj 72]. This

method involved coding the character in the form of a string

of T-tuples by tracing the character along pcints in it. An

appropraite set of T-tuples was chosen to trace the pattern

frcrr \ set of selected points, like an end point, junction

point, top left point, bottom left pcint, tcp right pcint

etc. The code represented by the sequence of T-tuples was

very compact. (Figure 2.27). A dictionary of codes for the

basic letters was prepared including information like the

CM. I.HiaHFp^JIMJOFH-1

la)

c-o, KJeMfoiioftjofH-i

cm iciH-crHei

Ul

cu rOirtstnroi

lay

CO. I0FHFM CW.:IOFMfDl

(g) 0>>

Kxiuii|>len illiMlruliiiK the im-tlio-curvo ctxliutc |iriM-etluic.

R^ave 2.- 2.^ Oo -"nS.-OLoo.c. ^---Uw.|xroOLeW_. l_R^'72.J

'tr

TELUGU ChAkACTtZ ^ECCGMHi^ P -. -; e 3 ^

length of the curve segments, etc. The code of the test

letter was compared with the codes in the dictionary, and a

SOX recognition rate was obtained.

It can be seen from the description given in section

2.4 that Telugu is a complicated but structured script which

has many curvilinear components. kecognition of similar

Indian scripts by [Sin 79], [Chi 80], [Cha 83], and

recognition of Telugu characters by Rajasekharan and

L eeksnatulu [saj 77], suggests that syntactic or structural

methods will be the most appropriate method for Telugu

character recognition. Vs i t h this view in mino, syntactic

nethods ana pattern languages in particular were examined

for this stuuy. It was observed that a pattern description

language, similar to PLANG, suggestec by [Sin B 3 J -culo be

required. However, such an implementation would be complex

and time consuming. Thus this thesis was directed towards

statistical methods. The details cf the relevant theories

anc algorithms employed in this thesis are described in

Chapter 3.

cFAPTER 3

RELEVANT THEORIES AND CONCEPTS

3.1 Relevant Theories And Conceptpt s

In this thesis, the decision theoretic or

classi f icatory approach was usee for recognition of Telugu

characters. Figure 3.1 describes the functions of the

recognition system, which basically consists of two parts,

dnalysib and classification. t>nelysis consists of feature

selection using projection profiles and cross-sections, and

adaptive learning to train the system. Classification

consists of feature extraction and Discrimination using a

similarity fucntion.

This chapter contains a detailed description of the

various methods of feature extraction, adaptive learring and

classification functions used in this study.

3.2 FEATURE SELECTICN AND EXTRACTIUN

The following methods of feature selection were usee in

this study:

1 : Application of FFT to the projection profiles

H Cross-sections or column and rcw runs

j: Application of FFT to the cross-sections

?

'J. .

<_

0-J

3CI

...

rc

<_

_.

2

0

F<i

-I

lil

co

-i

_c

i*

LuJ

o P

2 "-

Ui

1

_5_

\

-> L>

0

0

a.

2. ALGORITHMS

FURORE 3*2.

feature Extraction -

k\e~thod z

POURIER. t__=SCft\PTOROP"

Pft03"ECT)00 PROFILES.

I. VE-RITCAL PC-OTECTIOO PROFILE

1 II

I ,

11

i >

1

1 1

1. <

. ' 1

I 1

*

V 4 ,

N

\>so*co t_r p(ip

2- HOft)^c3K_Tie\L_ PROCTEC-TlOW p^Q>J= LE

i i . ,i a

- - f

> - -

3- pp_^Tc>iA<S<_>rJAL Pt-iOaFCTiON pPOFILE-

LEFTM

A. R*^*4X DIAGoOOAL, ^ROTECTTOW PROFMLfc

v-i XI

S'e~,

TVI- fccKcc: / C-.E CfAlT^ED fCiR"TWO R^SOMS : -. "Wry jour co*n*w HOch ose*u-

IWPO

P1-.KED0P g7 OTOCR WAqtfWAL

TELUGU ChAkACTtf .ZClGMTIOn Pc.je 39

4: Condensed cross-sections

5s Condensed cross-sections after noise removal

The same methoas were applied for feature extraction also.

3.2.1 Projection Profiles -

The projection profiles of an image can be obtained by

summing up the grey values in a given direction. For a

binary image (black and white), this simplifies to counting

the number of clack dots (representee as l's) in a given row

or column. Peaks in projection profiles can indicate the

locations of najor Tarts of the object. Projection profiles

can be generated in cifferent orientations namely rows,

column, left diagonal and right diagonal. For Chinese

characters, Oevanagari script 3nd Tamil characters, which

have more horizontal and vertical strokes, the horizontal

anc vertical projections are more useful. It is not clear,

however, how useful these will te for Telugu characters

which consist of a lot of curves. It appears that the

diagonal projections could be more helpful in this case as

they woula contain information of the curvilinear

components.

Figure 3.2 explains d i agr am a t i ca I I y , the methocs of

oataining the projection profiles. For a given hx\ matrix,

there will oe IN vertical, N horizontal and 2 N ciagonal

projections. In case of diagonal projections, however, only

the center N lines were considered, i.e., the corner

triangles witn sides N/2 X n/2 were omitted. In particular,

TELUGU CHAk^CTER ZclGMTIl'n Pd^e 4 0

the top left and bottom right triangles were omitted for

right diagonal projection, and the bottom left and tcp right

triangles were omitted for the left diagonal projection.

This reduction was done for the following reasons:

1. The size of diagonal projections would be N, for

an iNxN matrix instead of 2N.

2. Most of the useful information is usually contained

at the center of the picture, hence the corners can

oe cmrritted.

3. *Jhen both projections are considered, the corners

omitted by the right diagonal are included in the

left

diagonal and vice-versa, without loss of information.

[he algorithms for obtaining projection profiles in

various directions are described below:

For a binary pattern matrix PAT(NXN):

HORIZONTAL: HSUfU i ) = PAT( i , j ) , i = 1,N

VERTICAL: VSUM( i ) = PAT ( j, i ) , i = 1 ,N

LEFT DIAGONAL: LEFT(i) = PAT(i,j), k = (j-i)+f\/2

RIGHT (JIAGO'n'AL: RIGHT(i) = PAT(i,j), = <j*iMN/2

tFVk 3-3 .PPuj-ecTICnjpwOFH-E-. C-iV'G)- Fi'K cz/

o**<

-0xji1"iii11i222222 22223333 3333 53444^4 44^44

1234b67BS012343fi7 8SOl23 4b67 8',90l234Dfi7a90i2i4i675sOi23'ib678^0i?34S 0000OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOoOoOvOi/OOOOOuGoOuOOooooouooo 0

6 GOOOoOOOOOoOOOGOOOOOoOOOOOOOOOOOOOoOoOOOCOGOOOQOQOuOOGOOOOOOoOoO o

7 OOOOCOOOCOOOOOoOwOcOOOOOOOOOOOOnoOOOOOOOOOOOOOOOuOoOOOviOoOOOuOOO 0

8 OOOO OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO 00GOCOoOOOuOODiiOOOuC 0

9 OOOOOOOOOOOO OOOOOOOOOOOOOOOO JO OOOOOOOOOOuOOOoOuOuOoOuOonoOOOoOOO 0

10 OOOOGOOOOOOOGOOOOOOOOOGOQOi -OlOOOOOOoOoOOOOOGDOOwOoOoOOOOOuOOOGO<u 2

11 OOOOOOOOOOOOOOOOOOOOOOOOOtllUlOG 00OOOOOOOOOOOOOOOOOOOO OuOOOOOOO 6

"

12 OOOOOOOOOOOOoOCOOOGOOOOOiUllUHOOOoOuOoOOOlllliOoOGOOOOOOOoGOO \A

13 OOOOoOOOOOOOOOOOOOOOOUllllllliUOoOOOoOuOlUliliUOoOOOOOoOoOOO 2.)

14 OOOOOOOOGOOOOOOOOOoOollllUlOU 111OOOOOOoOiUliUliloOuOOOOOoOOO 2.2.

15 OOOOOOOOoOOOOOGOOOCOGUlOOOOOOlUlOOOOoOollloOOOllHlOoOoOOOOOOO i5

16 OOOOoOOOOOGOOOGOOOGOlUOOOOOOOOUllOOOOOollOoOOOOllHOuOGOOOGOOO ,3

17 OOOOO OOOOOOO OOOOOO 001 llOOOOOoOOOillOOOoOlUOOOOOOOUi J OOOOOOuOoO 13

18 OOOOOOGOOOOOOOOOOOOllllOOOOOOOOOlliOOOOlllOOOOOOoOH 11 oooooooooo \4

19 OOOOOOOOOOOOuOOOOOOUllOOOOOOOOOlliOOOOlllOOOOoOoOolll OOOOOOOOOO i2>

20 OOOOOOOOOOOOOOOOOOOUliO OOOOOOOO iliOoOulilOOOOOCOOGOUG 00 OGOGOOO )2_

21 OOOOOOOOOOOOOOOOOOlllllOOOGOOOOOlllOuOOlilOOwOOOwOGOltlOoOoOuOOO j^

22 OOOOOOO OOOOOOOOOoOl 11 11 loOOOOOOOUOOOOOlllOOOOOOuOOOillOOOGOOOOO iA

23 OOOOOO OOOOOOOOOOOO 1 1111 1 0 0000001 i 0000001 11 OOOOOOOOOO IUOOO0O0O0O K

24 OOOOOOO OOOOOOOOOOO 1 101 11 OOOOOOO 1 1 OOOOOO 11 1 00 OOOOOOuOlllGoOoOoOGf1 I'D

2b OOOOOOO OOOOOOOOOOO 1 101Hi 00001 11 00000001 1 1 OOOOOOOOOO 1 1 iOoOOOoOOO 1-5

26 OOOOGOGOGOOOOOGOOOllOlilllllllllOOOOGOollloOuOOOoOoOOJlOOnoOOOQO js,

27 00 OOOOOOOOOO 00 oOOlllGOlllllllllOOOoOuOOUlOOOOOOOOoOOl 100 0000000 17

28 OOOOOOOOOOOOOOOOOOllOOillHlllOOOCOOoOJllllOoOGOOOoOO] lOuOuOOOOO 16

29 OOOOOOOOOOOOOOOOoOllOOOOlllllOOOOCOOyOOliliOOOOOoOoOlllOoOoOOOOO H

30 00000000000000000011 00 OOOOOOO 000 000OuOOOl 111 OOOOOOOO UiOOOoOOOOO =)

31 OOOOoOOOOOOOOOGOOOHOOOOOOOOOOullllllllliUllOOOOOoOiUOOOOOoOoO 17

32 ooooooo 00 00 nooo 00on 000 000 000 001 111 llllll 1 1110000 Oo 01 11 00 OOOOOOO 17

33 OOOOOOOOOOOOoOOOOOilOOOOOOOOOOOliUllllllUOOOGOoOoOllOriyOOOOOOO 16

34 OOOOOOOOOOOnOOOOOOlllOO OOOOOOOOOOOOOOOtJOoOoOoOOOOOOOilOOOOOOOOGO 5

35 OOOOOOoOOOOOOOOOOOlllOOOOOOOOO'jOOOOOOOOOoOOOOOOOuOOOilOOOOOOOOOO <5

36 0O0OOOGOCO0O00GO0O111O OOOOOOOOOOO 00OoOGOoOoOOOGOGOGli 1 uOOOGOGOGO :.

37 OOOOOOoOOOOOOOoOGOOlilOOOCOOOOOOOOOOOOOOoOOOOOOOOOGliloOOOOOOOOO 6

38 00OOOOOOOOOO OOOOOOolllOOOOOO 00 OOOOOO JOOOuOoOOOOOoOiUOOOoOoOOOoO 6

39 OOOOoOOOOOOOOOOOOOOOlllOOOOOoOjOOOOOoOOOoOJOjOOOOOHiOOOOOoOOOOO G

4,0 OOOOOOOOOOOOOOOOOO 00111100000 OJCOOOOOOJOOOOOO 00OOOHOOOOOOOOOOOO 6

41 OOOOOOOOOOOOOOoOOOOOillllOOOOOOOOOJOOOOOOOuOOOOOolllOOoOOOOOoOGO ?

42 OCOOOOOOOOoOOOOOOOOOOllllOOOoCoOoOOOoOuOOOoOOOuOOliOO^OOOOOOOOuO 6

43 OOOOOOOOOOOOOOOOGOCOOOllillOjOOOO'OjOOOOOOOOOOOOOlloOOooooooooooo 7

44 00 00OOOOO 000000000000001 HllOOOOoOOOjOOO uOOOJOOliOoOOOoOOOoOuOOO 7

45 OOOOOOOOJOOOCOOOOOOOOOOOlUHO OOOOOOOOOOOOOO0OIUO0O oooooooooooo Z

46 OOOOOOOOOOOOCOOOOOOOOOOOOllllllOOOOOoOOOOOOOlUloOuOOOuOOOoOOOuO ic

47 OOOOOOOOJOCOOOOOOOOOOOGOJOlUlilllOOuOuGGOiUllOuOGOjOoOoOOOoOGO 13

4b OOOOOOOOOOOOGOGOOOOOCOOOjOOlllillllllUllllliOOOOOGOGOGCuOwOOOOO |g

49 OOOOOOOOOOOOuOOOOOOOOOOOOOOOOOOlllillllUliOOOOOijOoOuOOOOOOOoOuO i2

50 OOOOOOoOOOOOOOOOGOOOGOGOOOOOOGOlllllllll-lOOGOGOoOoOuOOOOOoOOOOO 11

51 OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO oOoOjOOOOOoOOOOOyOOOOOjO o

52 OOOOoOOOOOOOoOOOOOOOOOOOOOoOOOoOOOO/)OOoOuOoOOOuOoOOOoOuOi.'>OOoOOO c

53 OOOOOOOOOOOOOOCOOOOOOOOOoOOOOOoOOOOOoOoOOOoOoOuOoOOOOOOOOOOOuOoO o

54 oOOOoOOOOOOOOOOOCOOOOOOOOOOOOOOOJOOOJOOOjOiiOGOuOoOoOGOuOoOoOonyo c

55 OOOOoOOOOOOOuOOOoOOOOOOOOOOOOOOCOOOOOOuOOOOOOOoOoOoOwOO OOOOOOOOO O

56 OOOOOOoOOOOOOOOOuOOOOOOOoOOOOOOOOOOOjOOOuOuOuOOOoOoOuOuOOO&OuOOOC

-ooQO.erCzcoCcc.oO^rt^S-^^ .(_v.lv -v.

TcLUGU CHA,* ACTE:-'.<EClGM T I C \ F a ci e 41

Figure 3.4 shows the projection profile values for

character 'gp*. As it can be seen, the prcjections in the

horizontal direction are nothing out the total number of l's

in that row. For example, the projection for row 1G is 3

ano for row 11 is 6. Vertical projections for columns 18,

19 and 2o are i, 16 and 12 respectively.

3.2.2 Cross-sections Or Runs: -

Just like projections, c r o ss-sec t i on s can give an

inoication of the characteristics of the pattern. For a

binary image, where black dots are represented by l's and

white spaces represented by O's, a cross-section in a given

cirection is given by the number of distinct sequences or

runs of l's.

The cross-sections are computeo using the following

algorithm:

For row runs:

FOR I = 1 TO kCWS

ROfcRUN(I) = G

FOR J = 1 TO COLUMNS

IF PATTERN( I , J ) = 1 THE in

IF PATTEkN( 1-1 , J) = 0 THEN

START * \E *UI\

kUWRUN(I) = kJV*RON(I> + 1

The column runs can be obtained in a similar way.

TELUGU ChAKACTtr! << t C C GM T I l ', P :_ 4 2

The number of column runs wilt te equal to the number

of columns and the number of row runs will be equal to the

number of rows. This basic set of feature measurements will

be unique for each character. For the current study, the

first and last four run values were omitted as it was

observed that they were mostly zero's, obtaining a reduction

in number of features from 64 to 56.

The row and column cross section values for character @P

are shown in Figure 3.4. In row IC, there are two runs of

l's, one starting at column 27 and the other starting at 30.

The length of the run is not considered here; cnly the

number of runs is important. In ro- 11, there is only one

run of l's starting in column 26. Similarly, in column 20

there is a long but single run of l's, starting at row 18.

Column 29 has three runs starting from 11, 2c and 45.

The column run for this character after deleting the leading

and trailing C's is :

1112223333333343332222 32 23332222 22 2211

and the row run is:

21223444444 44 4554444333322222222222222L11

Typically, the row and column run values jre represented

witnout leading and trailing u's.

3.2.3 Concensed Cross-sections-

\-\Cr 3.4"

CQLUMtf AND ROW CRCiSS-SECTIOrt VALUES FOR TtiLUGO CHaP-ACIKk 'AA'

Illllllll4.222222222233333333334444444444555b5b5b5b66666

1 234567 8 90 I234b67 8 90123456 7 89012 34567 8 90 l 2345678*01 234567 8*0 1234

5 ooqoooooqoooqoooqoqoqoooooqoqoqoooqooooooogooooooOooooqogoqooooo 0 P

6 oooooooooooooooooooooooooooooooooooooooooooooooooouooooooooooooo 0 Z7 oooooooocooooooooooooooooooooooooooooooooooooooooooooooooooooooo 0 P-

8 oooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooc 0 ^9 oooooooooooo oo ooo oooooo OOOOOOO 000 00 OOOOOOOOOOOOOOOOOOOOOOOOOOOOO 0

10 OOOOGOOOOOOOOOOOOOOOOOOOQOHOIOOOOOOOOOOOOOOGOOOOOOOOOOOOOOOOOOO 2

11 00000000000000000000000001 HIHOOOOOOOOOOOOOOOOOOOOOOOOOGOQOOOOO 1

12 OOOOOOOOOOOOOOOOOOOOOOOOlllllllllOOOoOOOOOoOlllllOoOOOOOOOOOOOOO 2

13 00000000000000000000011 111 lllliUOQOOOoOOOlHllUUOoOOOQOoOOOOO 7

14 OOOOOOOOCOOOOOOOOOoOolllllllOUlllOOOOOOoOililililllOOOOOOOOOOOO 3

15 00000000000000000000011 lOOOOQOllllOOOOOOolllOOOOlllUOGOOOOOGOOO 4

16 0000000000000000000011 10000000011 llOOOOOollOoOQOOUHOOOGOOOGOOO 4

17 OOOOOOOOOOOOOOOOOOOO 1 11 OOOOOOOOO! nooooomooooooo ll 11 OOOOOOOOOO 4

18 OOOOOOOOOOOOOOOOOOOllllOOOOOOOOOlllOOOOlllOOOOOOoOHil OOOOOOOOOO 4

19 OOOOOOOO 000 OOOOOOOO 1 11 10000000001 lioooomooooooooouioooooooooo 4

20 00000000000000000001111000000000111 OOOOUIOOOOOOOOQOIIGOOOOOGOOO 4

21 0000000000000000001U110000000001 1100001 HOOOOOOoOOOlllOOOoOoOOO 4

22 OOOOOOOOOOOOOOOOOOllilllOOOOOOOOllOOOOOlllOOOOOOyOoOil 1000000000 4

23 oooooooo oooooo OOOO 1 11 111 00000001 iOOOOOOlllOOOOOOOOoOlUOOOOOoOOO 4

24 OOOOOOOOOOOOOOOOOO 11 01 11 00000001 lOOOOOolllOOOOOOOOoOl 11 OOOOOOOOO 5

2b 0000000000000000001 101 11100001 11 000000011 100 JOOOOOOOIIIOOOOOOOOO 5

26 ooooooooooooooooooi 1011 111 11 11 110000000 llloOOOOOjOoOOll OOOOOOOOO 4

27 OOOOOOOOOOOOOOOOOlllOOlllllllllOOOJOOOOlilOOOOOOOOoOOl 1000000000 4

28 ooooooooooooooooooi 1 001 111111100000000011110000000000] 1000000000 4

29 OOOOOOOOOOOOOOOOOOI 10000111 llOOOOOOOuOOliHOOOOOOOoOl 11000000000 4

30 OOOOOOOOOOOOOOOOOOI lOOOOOOOOOOOOOOOOuOOOHilOOOOOOOOUiOOOoOoOOO 3

31 OOOOoOOOOOOOOOCOOOllOOOOOOOOOOOUllllllUllllOOOOOoOlliOOOOOOOOO 3

32 OOOOOOOOOOOOOOOOOOI 1000000000001 11 UlllllUllOOOoOoOlliOoOOOOOOO 3

33 oooooooo OOOOOOO 000 11 000 000000001 111 llllll HOOOOOuOoOl 10000000000 3

34 ooooooooooooooooooi UOOOOOOOOOOOOOOOoOOOOOOOOOOOOOOOx 10000000000 2

3b oooooooooooooooooomooooooooooooooooooooooooooooooonooooooooGO 7

36 oooooooooooooooooonioooooooooooooooooooooooooooooou 10000000000 7

37 OOOOOOOOOOOO 00000001 ll OOOOOOOOOOOOOOOOOOOOOOOOOOOOO ll 1000 ooo OOOO 2

38 oooooooooooooooooooinoooooooooooooooooooooooooooomooooouooooo 7

39 ooooooooooooooooooooiiioooooooooooooooooooooooooooiuoo ooooo oooo ?

40 QOQOQOOOGOOOOOOOQOOOIUIOOJOOOOOGOOOOOOOOOOOOOOOGOIIOOOOGOOOOOQO ?

41 OOOOOOOOOOOOOOOOOOOOiliUOOOOOOGOOJOOOOOOOvOOOOOolilOOOOO OOOOOOO 2

42 QOOOOOOOOOOOQOQOOOQOOllllOQOOOOOoOOOuOOOOOoOOOOOOllOGOOOOOOOOOOO 7

43 OOOOOOOOOOOOOOOOOOOOOOllillOOOOOOOuOOOOOOOOOOOOOHOOO OOOOOOOOOOO 2

44 OOOOOOOOOOOOOOOOOOOOOOOUIUOOOOO OOOOOOO y OOOO 00 HOoOOOOOOOOOoOOO 7.

45 oOOOOOOOOOOOOOOOOOOOOOOOlUHOOOOOOOOOOOOOOOoOUlOoOOOOOoOOOOOoO 7

46 oooooooooooooooooooooooooiiiiiioooooooooooooiiuoOGOoooooooooooo 2

47 000000 00000 0000000000000001 UlUllOOOOvJOjOHlliOoOGOOOoOGOOOGOOO 7

48 oOOOOOOOOOOOOOOOOOOOOOOOOOOlilillllUHllUllOOOOOoOGOOOgOuOOOOO 1

49 OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOlllillllUllOOOOOoOoOoOOOOOOOOOOO 1

50 00000000000000000000000000000001 llllllllilOOOOOOoOoOOOOOOOoOOOOO 1

51 00000000000000000000000000000000000 oo ooooooooooooooooooooooooooo 0

52 oooooooooooooooooooooooooooooooooooooooooooooooooooooooouooooooo 0

53 00000000000000000000000000000000000OOOOOOOOOOOOOOOOOOOOOOOOOOOOO 0

54 oooooooooooooooooooooooooooooooo JOOOJOOOjOuOOOOOoOoOGOOOoOOOOOoO 0

55 oooooooo ooooooooooooooooooooooooooooooooooooooooooo OOOOOOOOOOOOO 0

56 OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOwOQOoOQOOOO OOOOOOOOO 0

000-00000000001 11 2223333333343332222322333222222221 lOOOooOOGOuOOO

COLOViM RUN VALOR

TELUGU CHArACT.^-ZCCGMTZ'f- P _. g e 4 3

The condense cj column-run string is formed by retaining

only one digit in each run of that digit in the

corresponding column run. Tn i s can be cone by simply

scanning the string ana replacing a strirg of successive

duplicate entries by a single entry. The condensed row run

is obtained in a similar way.

From Figure 3.4, the full column run for character'^>'

i s

Ill22z33333333433322223223332222222211

anc its condensed run is

12343_32321.

This is obtained by replacing the first 111 ty 1, 222 by 2,

3J3333 3 3 by 3, <? oy 4 etc.

Similarly the row run is

21223444444444554444333322222222222222111

ano its condensed run is

2123454321.

The advantages of using condensed runs are:

(1). Condensing is a kino of thirrin j process.

(2). Trie condensed strings will be inaepencent of the

thicKness of the letters and tne proportions cf the

b i a n r y niuge.

(3). They require less storage space. ( r educ t i on from t4 to

TELUGU iHAkACTE'. ZcLGMTIC:, Page 44

abou t lc ) .

(4). The condensed run strings can be unique, qualifying

the rr as good features for reccgnition.

(5). Easy to compute.

(6). A simple string comparision function can be used for

classification.

3.2.4 Noise Removal In Condensed Strings -

'.oise removal algorithms can be applied after the

condensed runs are obtainea. The method usee in this thesis

is derived from Siromoney ISir 1 i ] and involves removing

solitary occurences of any numoers from the condensed runs,

since sucn occurances are often due to noise genearated in

digitization. A perfect example is shown in Figure 3.4 in

row 1 j at column 29. There is a'C'at this position -ina

the resulting gap is generated by noise. Because cf this,

now there are two runs of l's, 11 anc 1, though it should

have a single run of 1111. This can create variations in

cross-sections ana lead to rejection by a reccgnition

algorithm. This is a case of an isolateo occurarce of a

run. It can be seen that if the gap is filled, it will oe

consistent with the subsequent runs, but it is very

difficult to know whetner filling the gap would be useful or

culc create nore noise. It was chosen ir this thesis to

eliminate such occurances as it was very simple to

implement. Another example is at row 21 and column It.

There is a single 1, (or a blacn dot) in comparison to its

TELUGu ChAkACTtk i-ECCGNITI P 3 ^e 45

immediate neighoor column 1^, which has a nn of lergth 16.

This is obviously generated oy noise and should be

eliminated. This method of eliminating isolated occurances

of runs usually eliminates spurious features being

considered. However, there is a possibility that important

information could be lost.

Noise removal can be implementec with the help of a

count field associated with each value in the cencensed

cross-section. In the first step, the cencensed

cross-section is ootained along with the the number of times

a numeral occurs in the full run. In the second step, a new

string is formed by deleting all the elerrerts from the old

string with a count of 1. As a last step, the list is made

unique oy merging successive elements of the same value, and

.j doing up their counts.

From Figure 3.4, the full column run for character'@J'

i s

111222 333 3333343332222322 33 32222 222211.

Its condensed run is

12343232321

mc its count array is

j38 134 12 3 82.

This is obtained by replacing 111 oy 1, and setting the

count to 3, replacing l22 by 2 and setting the count tc 3,

TELUGU ChAPACTE; ; C G N I T I Page 46

repalcing 33333333 by 3 ana setting count to 8, etc. It can

be observed that there are two isolated occurances for 4 and

3 with a count value of 1. After removing these two

numerals, the run values are 123322321 with associated

counts 336342382. Now, successive runs of 33 can be merged

to a single 3, and runs of 2 2 to a single 2 by adding their

associated counts. The final runs values are: 1232321 with

counts 3386382, where'8'

is 11.

2.3 Magnitude Spectra Using Fourier Analysis

Numerical properties of projections, such as their

one-Gimensionsal moments (Fourier coefficients)* can be

useful as object descriptors. Fourier analysis is a general

purpose prooleir solving tool that uses a transforx technique

from space domain to frequency comain. Thi_ is of

particular significance to character recognition because the

magnitude spectra of a character will be in frequency

domains ana are independent of the position of the

cnaracters. That means that the character can be shifted

(but not rotated), and the Fourier descriptors still will be

essentially the same. This is sometimes referred to as

spa tial in variance.

The other advantages of using Fourier transforms (FT's) are

that they are simple to use ano they provide excellent cata

recuction. ,1anytransforms give N descriptors for N input

values. Cnly N / 2 values need oe consicerec because cf the

T E l U c jLhAMCIi..- -li.LvjMTI-N Page 47

symmetry property of the FT. Even further reduction can be

obtained by observing the frequency spctrum and eliminating

the less significant components (Zri 78], [Son 5], [Per

773 , [Sta 76],

The basic defintion of the Fourier transform operation is

UZc 7^J

r-lwac

where: F ( w ) = the frequency transform

r(x) = the function to oe transformed

w = the freouency v a r i n c I e (e.g., radians /second)

x = the spatial variable ( e . a . , time in seconcsl

j =sqrtof-l

Tne transfcrm F(w) is in the complex domain ana is oerfcrmed

using complex aluebra and integration. The aosolute value of

F(w) is defined as:

t-

!F(w)J = F * ( w ) F ( w )

To calculate JF(w),, we can use the interpretation

!F(w)Z = Ze F(i)}^* L I m F(w>}-

where

Re F(w = I^WG**-^-CO

ir f ( ) = r

j*^ s ivv Clfix>^

To finally arr^rje these equations into a structure

suitaole for computer calculations consider fi = f(xi) to

represent the ..\j13 value from an experiment or an equation

TELUGU CHAkACTZ ":CCGMTa.'< Face i8

at position xi. For simplicity we assume the data points to

be equally spaced, that is xi+1-

xi = delta x, a constant.

le can also consider fi to exist only over the interval

xl to xn. Outside that interval, fi will be defined to be

zero. Combining the above considerations into a

programmable form, we have:

i*-

The magnitude spectrum is simplylJF(wj)! .

This equation is the oasis for computer calculation of

the Fourier transform of the function representd by { f i > in

this thesis.

In this thesis, the Fourier transforms were use a -ith

projection profiles and cross-sections.

3.4 Learning Algorithm

A self adaptive learning algorithm was used in this

thesis to train the system, with oifferent sarrples of data so

that the various characteristics could be absorbed into the

system. This algorithm is a simplifiec version of that used

o y Uhr and Vossler [Uhr c3] which a. i justed the weights

associated with each feature and the values of the features

themselves. This method is independent of the values of the

features themselves which facilitated using the same

learning method for features extracted from different

TELUGU CHA-ACTt^ CC l G:\ I T I Z Faae h9

techniques. This met one of the objectives of the thesis,

to study various features ana see which one best describes

the character. In this study, learning was usea with

features obtained by Fourier transforms of projecton

profiles ana cross-sections.

The learning algorithm is as follows.

Each character has a set of feature measurements, called

typical values, stored in a dictionary. The number of these

typical values depenas on the methoa of feature extraction.

For each character there is a weight associated with each

typical value. Initially, all the weight factors are

initialized to 3 constant value (say bC). The choice of the

initial weight factor is important because its value should

be neither too low nor too high after learning. The

following steps are followed for each feature set to adjust

the typical values and weights associated with them.

TELUGU CHARACTER kZGGMTIlN ^j.e !

1. Initialize the weight matrix

2. Input one sample character per standard character to

the system.

3. For each input character, adjust the weights of the

features

of the appropriate characters by comparing the

sensitivity or ability of the features to differentiate

different characters.

For the kth character, then:

FCR J = 1 TO NUfFEATURES

FUR I = 1 TO NUC STANOARCCHAPACTEkS

IF TYPIcalVALUSa*PlE( J) - T Y P IC AL VALlE S TO ( K , J ) <

TYPICALVALUtSAMPLEf J) - T Y P I C AL V ALU L S TO ( I , J ) THEN

THE CORRECT STANOASD CHARACTER IS CLOSER TC THE

SAMPLE, SO PEwARC IT

WEIGHT(K,J) = W6IGHT(K,J) * GPTIMUM

nEIGhT(I-J) = HEIGhT(I,J) ? 0PTIMUP2

ELSE

THE CORRECT STANUARO ChARACTER IS NUT CLCSER

TO THE SAMPLE, SO PUNISH IT

hEIGHT(K,J) = WE1GHT(K,J) - CPTItfUfl

WEIGHT(I,J) = wEIGHT(I,J)- CPTIMUrZ

The values of CPTJ.iMU.Ml AN 9 OPT I "UK 2 are choosen in such

a way that G P T 1 1* U fi 1 > 3*0PTI:VIU.V,2. This selection nasically

prcvides a disctinction between important anc less irtpcrtant

features.

TELUGo CHArACTr--.tCCGMTICN Page 51

At this point, after each learning phase, the system

can be tested and the performance of recognition and the

effect of learning can be studied.

4. The typical value is also adjusted to the average value

of all the samples so far. A COUNT is maintained, one for

each feature set. Its value is set to one at the time the

dictionary is prepared. After every learning phase,

its value is bumped by 1.

UCUNT = CCUNT * 1

TYPICALVmLUESTD =

( TYPICAL VALUESACPLE + ( TYP I CAL VALUES TC )*( COUNT-1 ) )

/ CCUNT

5. This learning process can be repeated for all

available sets of samples. Ideally, the larger the

numoer of samples, the oetter the chances of correct

recognition.

3.5 Classification:

The classification or assignment of the unknown input

to a particular pattern class is accomplished by using a

difference polynomial as a measure of similarity. This

method was based on the polynomial used by Zles Zil 3 u ]

for pattern analysis and involves the following steos:

1. Normalize weights.

normal izedweight = rawweight/averageweight

TELUGU CFA(vJt_TE- M T I - N Page 52

where averageweight = sum of weights/number of terms

2. Compute the difference metric

totdif = sum of normal izedweight * cifference >

where difference =

! typ i ca I va I ues td -

typ i ca 1 va I ue te s t !

When both patterns are identical, totdif will be zero,

otherwise its value hopefully will be minimal for the

correct standard character. Its value can te adjusted to

fall in the range of 0 to 1 if the typical values arc their

differences are normalized. These values also need to be

normalized if the r a n 3 e of typical values is high. In

eitner case, the difference metric with minimum value will

hopefully be the correct character, the next higher value

will be one tor the next closest character etc. Also, it is

necessary to normalize the weights, as they vary a lot

during adaptive learning.

This difference polynomial was used in this thesis with

the fourier descriptors of projection profiles and

cross-sections.

A simple string comparision function was usee with ccnoensed

cross-sections and cross-sections after noise removal.

The results of various metnoos of feature extraction anc the

atfact of adaptive learning eescrioec here wilt be discussed

in Chapter 5 .

CHAPTER 4

IMPLEMENTATION AND TEST PROCEDURES

4.1 Implementation And Test Procedures

The fucntional block diagram of the telugu character

recognition system is shon in Figure 3.1. The system

oasically consists of three parts: pattern analysis,

learning and recognition.

This chapter contains the description of the hardware

configuration usea and the software packages uevelcped for

tne study. The details of test procedures are also

oescr i bed.

4.2 Hardware Configuration

This study used two different hardware configurations:

an I8M PC for digitizing the characters and a VAX 11/78G for

performing the functions of pattern analysis and

recognition. The VAX computer system was selected so that

complex analysis and recogniton stucies could be easily

performed with large data sets.

A oi g i tal viaeo camera interfaced to an I <i "i PC was used

to digitize images of the characters. The ci^itizec imajes

w= r e stjrea on a floppy disk and transported to a VAX where

the pattern analysis and recognition functions were

T E L U G U t_i-iARACTE- K E L C G N T I u N 55

performed. Information aocut some of the availacle digital

video camera equipment can oe found in [Cia-1 83J, (Cia-2

831.

Figure 4.1 shows the details of the hardware

configuration. The equipment usee consisted of a cigital

viceo camera made oy SCNY anc a vioec monitor interfaced to

the IBM PC through an HS232 port. A powerful software

package supplied by I M A G E L A 3 was used to digitize the

characters. The charcters were written on 8 .5X11 inch

paper, 16 characters per oage, ano placed on a flat

backlignted toara. Another light source was projectd

cirectly from tne top. The picture was purposefully

overexposeo to get maximum contrast. A little screen on the

viceo camera was used to adjust the paper within the frame

of the video monitor screen. Figures 4.3 s h o s the set of

standard characters and the sample characters used f cr tnis

study can be founa in Figures 4.3 - 4.11. These samples of

handprinted characters were collected from three different

peop I e.

Once the image was digitizea anc frozen, several other

functions were performed. The image was edited to eliminate

any apparant noise around the characters. The grey scales

tnen were adjusted to get maximum contrast and to nc realize

the values to C or 256,with'J*

representing a white space

ana'256'

representing a black spot. This was necessary in

order to eliminate noise generated ace to the texture and

uneven ness of the paper. Next, the image was zoorrec to a

!a.

o

&2

8

2A

I

_J

C_0

TElUuU CHAkaCTE^ f- LClZT [UN race 5 *->

256X256 pixel size from the original 512X512 size, to get a

nice size of 64X64 bytes for each character. Finally, the

image was archiveo on the disk. Each page, containing 16

characters, was stored in a single unformatted file of 64k

oy t es .

Both standard characters and samples for learning were

digitized under the same conditions.

A second ISP PC connected to the VAX * a s used tc unload

tne image files. All the 6-fk byte binary unformatted files

cn the floppy were converted to VPS files of 129 records

with a fixed record length of 510 bytes. Format conversion

routines were built to extract the binary images of

incividual characters form these files.

4.3 Software Configuration

The software for analysis, learring and recognition was

implemented in FORTRAN 77. FURTRAN is a nice language for

scientific, and mathematical applications anc was selected

as the programming language, because it was easy to

implement the Fourier Transform functions. Tne EQUIVALENCE

statement in FORTRAN also came in handy for file format

conversions.

The software configuration of the character reccgnition

syteir is explained in the dataflow diagram in Figure 4.2.

Tht: circles represent software functional rrocules with lata

flowing in and out of them. Oata files are denoted oenea th

I

Iu

_

ivj

rl

TELULU CHA* AC TEF-, -vCcCGf. TI Pae 57

horizontal lines. In adcition to several test programs,

which were developed during the testing phase, the final

software basically contained five programs: DICTIONARY,

SAMPLE, LEARN, RECOGNIZE ANC CRUN. A list cf the functions

of the programs is given below:

DICTIONARY:

Prepares the dictionary of standard characters with

features extracted from the binary image file.

Reads the binary image file I 129 records cf 510 oytes).

For each of the 16 standard characters, computes

hor i zon ta I ,

vertical, and diagonal projections and their fourier

coefficients, computes column and row c r o ss-sec t i on s ,

and writes one record to per character to std.dat file.

SAPPLE:

Prepares sample data file with features extracted from

different samples for each stancard character.

r.eads the binary image file (129 records cf 510 oytes).

For each of 128 sample characters, computes horizontal,

vertical* and diagonal projections and tneir fourier

coefficients, computes column anc row cross-sections,

and writes one record to sample.dat file.

LEARN:

Uses the self adaptive learning method to adjust the

weights of the features of the standard characters to

absorb the characteristics of the sample characters.

Reads data from std.dat file and updates it accorcmy

TELUGU CHARACTER KtCCZTIUN_

5 a

to the values read from sample.dat file.

RECOGNIZE:

Classifies the test pattern accorcing to the dictionary

using a normalized weighted difference polynomial. The

correct character will have the minimum value of this

polynomial. The values of the polynomial for each

of tne six features is printed for each standard

in 6 for all i^ test characters.

cHN:

Lbt._ins cross sections of rows e n c columns, condensed

cross-iccti n s ana u_ _ I i e s noise removal c r o c

-

u u r e .

I e -, t

A smalt scoset of 16 characters from well over 2 0 u 0 of

the Telugu characters was cnosen for this study- The set

included a variety of characters which are similar out

distinct. They were selected in such a way that some of

.n i- m j h o u i _ j

-

<sy tor r ? c > g a i t i o n :: y ;_ r o j . : c T i n r f i I '

-.

anc some by taking cross-section^. h i 3 u r e h . E> bro-. tne

dictionary characters and Figures 4.3 -

4 . 1 1 s n c *

handwriting Sjmpt.es or t n r a a d i f t -

1- r r ,^'ccl?, T . 1 - st n c. -. r

characters are numbered from 1 to 16 and the samples are

numbered from 1 to 12?. The characters '^'.',-s', '_-','^'

are

all very similar and are confusing even to human oein^s.

Similarly, the characters'gj'

and'^p1

differ only at the top

right corners but are different ovels. The character '7V is

included as it is totally different from rest of the 15

Tl)\t.

a

a

Oi<Ca

Oft

u.

o)

TELUCU CHARACTtr ZCCGNTZN c a c e 5 9

characters. This data set, thus, is a small but

representative sample of most Telugu characters in that

there are both easy ano hard characters to recognize.

4.5 Test Procedures

A series of tests were performed to analyse the

recognition rte of the system with different methoas of

feature extraction and classification methods.

As discussed above, a dictionary of 16 standard

characters * a s prepared ana stored in a data file. A set of

_ samples for each stanaard character were also digitized.

Features were extracted using Fourier analysis of projection

prcfiles and cross sections.

The recognition was tested first without training from

samples.

a weighted linear difference polynomial, also called the

cityolock'

method (Did 7o], [cJI 80] was usee ds the

discrimination function. The same 123 samples which were

used for training the system were given as the test c a t a .

The next step in testing was to train the system with

cifferent handwriting samples. The learning program was run

to train the system with one set of samples, one fcr each

stancard character, and the classification was tested. The

classification program printeu the values cf the

discrimination function. The results of classification were

TELUGU CHARACTER <ECLGNTIGN Page 1 0

tabulated in a 'confusion matrix'

which shows the systems

ability to discriminate different characters. The above

steps were repeated 8 times, in oraer for the system to

absorb characteristics of all the samples.

The results of these experiments are described in chapter

5.1

Another series cf tests were performed using cencensed

cross sections as the feature extraction method. In tnis

rretnod, the column and row run values were prepared for both

standard and sample characters. The discrimination function

was a simple string comparision function to compare the

values of the dictionary and test characters. The results

are printed in Figure 4.14. The confusion matrix was

prepared manually. The same test was repeated after

"scrubbing"the data, le. eliminating solitary occurances

of single runs. These results were also tabulated in a

confusion matrix.

CFAPTEK 5

RESULTS AND DISCUSSION

The objective of this thesis was to study various

methods of feature extraction for recognition of cursive

Telugu script. This chapter presents the discussion on the

results of various experiments. Topics for further research

are suggested in Chapter 6.

The binary image of each character was represented by

6 4X64 bytes. Each cf the projection profiles is representd

by o4 oytes, and the power spectra of the projection

profiles also have the same size. However, using the

symmetry propery of the Fourier transform, only first 32

values of the Fourier descriptors need be considerec. This

resulted in a compact representation of the feature set. In

aocition, from an examination of the magnitude spectra, it

coulo be seen that the lower frequency components are

dominant, and that the high frequency components are

insignificant. This fact could be used for further cata

compression of features but was not considered in this

study. In all, a total of 123 oytes were recuired to store

-all four sets cf features for a character.

The values of horizontal, vertical, left diagonal and

right diagonal projection profiles and their power spectra

are snown in Figure 5.2.

TELUGU CnAfACTt^ SEC LGMTICN Page 6 3

The column ana row runs (cross-sections) were

representee by 64 bytes each. Frcm an observation of the

binary image, it was found that the outermost four lines in

both vertical and horizontal directions did not contain any

information, so without losing significant data, elements 5

through 5 o were selected as the features. This required 112

oytes of space for the two features. In later tests,

however, even further data reduction was achieved by taking

the power spectrum of the cross-sections and usino only the

most significant 32 values. Condensed runs required fewer

tnan 32 bytes, and further condensation resulted in even

fewer oytes.

The actual values of row and column cross-sections are

shewn in Figure 5.3 and their fourier desciptors are also

shown in Figure 5.3.

The dictionary of standard characters was prepared and

stored in a data file. Features of sample characters were

extracted in a similar way and the data stored in the

sarrple.dat file. The sample character set used for learning

was also used as test input.

The classification program was run with these test

samples, and the results are tabulate.-) in what is called a

contusion matrix'. In this matrix, rows represent the

input character and columns represent the recognized

character. Thus the diagonal entries correspond to exact

matches. These entries would have a maximum value cf o, as

TELUGU ChAxACIc" r h c C G N I T 1 C ; Page t- 4

8 samples were used for testing each of the 16 standard

characters. A value Of <8 indicates incorrect

classification, and the mistaken character can be found in

the same row.

The recognition rate was calculated as the ratio cf the

sum of the diagonal elements to the total number of samples.

A summary of various test results is presented in Figure

5.1.

5. 1 Resui ts Of Test 1

The magnitude spectra of horizontal, vertical, left

diagonal and right diagonal projection profiles were used as

features in this test, and a recognition rate of 64/. was

ootained. Some of the characters like,/_^'

,,r4'

,** '__>'

,

which are confusing even to humans, were recognized

correctly by the system. The confusion matrix for this test

is given in Figure. 5.4

The system performed very well after the learning

phase, recognizing 96% of the characters. This suggests

that the power spectra of projection profiles can be used as

a good feature selection method fcr a cursive script like

Telugu. However, in the case of the cnaracter',v

', the

system performed worse than its performance before learning.

This can be attributed to the fact that the samples are so

different that the original characteristics were acversely

af fected.

RGiU&ET 51

SU(^M^F-V QF . TFST gSgjLTS

resr*r FUTURE CLAW\F\t^T10N Lt^RN>lN(_r J5etOGcA)lT10r-

EXTR-Ac-TlorO (METHOD RATE

METHOD <%-

. PRccrECTiow pRofMlE. LtMEWJ->(FFER.eC_"

rye 64 /

foori fr rZ-esasirtofs POC1OK\#\-- L ^_fe

-) >? "__fe >ES 96/.

2l- O50S.S. -SECTIONS

POC<NJOfA.\Al-

NO 2.0%

; 5? 7PS2.31/.

3- FDDPIER DE-XPiTrofcS -> MO S2/.

V S) NS 9^/-

4.COLUVVM

con>i>ek_;q_ _?**

crd_c-_:etnolos.

STRVWGc to^p^fe\0iU too 3_5y.

>> tee* >) VO &>6>y.

>

eoN>D6>J-.eD Cousmkj

<1RCS-Semofos

" wo 2.5y.

*SC*U)l WO is-y.

<$> Si -0 **"

, 6 * i .ri

c; aos^a*

eO

i

A

I *

eo

"8'

/ 8

- *

/

_Z

/-

V, '/

- - - --

-

/

/ I

/1

%3

'

5

6 /2..-

S

8

i 7

s

3 /

8 /1 1 /

%\ /.

i%6 ,

?

s

r

7

7

I

V.1 ./

/i

i

|V/8

8

8

4.

/ %4

.

1/

V v, '"I*

Be***AFTEj^.

X or correctaeseiW** TV/*-* <

64- y.

y. f. coarserRetft^ms/o

_r\/l2& - **6 /.

5-4-

TELUGU crAkai_TEv'

<ZlCGUTIC-\ P a e 6 5

5.2 Results Of Test 2

In this test, the cross-sections of rows ana columns

were used as features, ana a linear difference polynomial

was used for classification. As can be seen from the Figure

5.2, the values are very close to each other, naking

discrimination difficult. Only 20'/. cf the characters were

classified correctly, and the performance never seemed to

improve even after learning. A 20% recognition rate

ooviosly suggests that tne features are net very aistinct

ano unique.

A major contribution to the failure of this method is

the spatial variance of the characters, i.e., physical

displacement of the characters within a given frame. In the

Subsequent tests an attempt was mace to eliminate spatial

dependencies of the features of the characters.

5.3 Results Of Test 3

The magintude spectra of the column and row run values

were used as features along itn a discrimination function

method similar to the one used in Test 1. The magintude

spectra yielded unique features ana an excellent nit rate of

co/. uefore learning and 9 9 7. after learning. The Fcurier

transform eliminated the problem cf spatial depencencies,

while cross-sections provided unique features, hence the

success ot this method. See Figure 5.6 for the confusion

Ci) <_? <S? <_> rf i4 z_ rf S3 Q^>, 3D 2d .

(

CDZZ

1 11

1

\ 1 1 1 i' 1

ry 1'

11

G** \n/ 6/_r "_ \ \ |

2-2 X i i n

oO

v\

1i

i i

i i ii

^ ! ^ i

?C yi

r

Xi

i ^ <z/ i

i

6 ^i

i

*-

2wA!

4 y\y

V6*=- n

&

yQJ

i

i

*/

Zy

. ,

aoL-W

2Q

r....

...

i ~7/

//

\~pK 1 1 I 1/

RATE--

Op (^-TCOG^vJlTK-VO

After le^ra_mog- c

1C_o/12& = *$-4~-

C %

FT* 5,7

gpQSS-ggi-.T10J - <TOLtJMK>

GD Sf SI oO 4 < t% r6 OO ej> a? S> #

CO ^ 1

G33

Xi

^ X.=0 /s'

yrt

-A

-/-

& yi

yr*

, y_5

yo &/

ao yS3 y&

,. . <u

K^-.nnoO SAT*-- *l/.M - "3*03Z*f*

W*^

TELUGU LnAixACTEiV ZCCGMT1CN Page 66

matrix.

5.4 Resul ts Of Test 4

This was an improvement over test 2, in that the

conoensed values of column and row runs were used. This

method was used successfully by Siromoney ISir 78] for

recognition of printed Tamil characters and was was expected

to yield oetter results because the condensed runs have the

effect of thinning and characterizing the features. The

discrimination method was a simple string comparison. On

the average, column runs Droved to be more useful than the

row runs. The standard characters had unique values,

suggesting that this method coulo te successful. however,

the values of the sample runs appear to be affected greatly

oy the length of the curve, curvature of the curve and

relative position of the different segments of the curves in

the character. Also, after a close examination of the

values, it was found that there were a number of isolated

occurances of single runs, i.e., a feature occur ing only

once. In genaral this can be considered to be noise,

although it might contain useful features. Improvement can

be made by obtaining symbolic runs, removing noise, etc.

Figures 5.7 and 5 _provide a comparison of now

condensed row and column cross-sections scored before and

after noise removal.

R-i^-8-

CD G? -.a eO -T *K 4 rf rf oO e><-S 4_^

GD1

i

G /_Q y-=0

^ y<

3/

4"

Xdv

z7

/rf

-

Xrf X< y

--

e;z4

SO Xya

?y

2_ /3

2-3/l.iS = 1* y' AFTEG NyOVSEP11250^710'0

TElUGU CHAKACTEk "cCCGMTIl,: Page 67

5.5 Resu Its Of Test 5

Noise was reduced for this test by eliminating isolated

occurances of single runs ot cross-section values. This was

done at the feature extraction stage, and the method is

described in section 3.2.4.

As it was observea in this test, the cisadvantage of

this noise reduction as that one may lose useful

information. This was observed to be the case where the

characters had a '/^ type of joint between segments.

This method yielded very poor recognition rates of 82.

No learning was attempted in this test, out training the

system with different samples might improve the reccgnition

rate.

Chapter 6

CONCLUSIONS AND FURThEk kESEARCH

6.1 CONCLUSIONS

This chapter describes the advantages of the various

character recognition schemes tested ir this thesis.

Suggestions on expansion and further o r k in t h i _, area will

also oe presented.

A character recognition system witn two different

methods of feature selection has been implemented anc tested

for Telugu character recognition. In the first method, the

projection profiles of the pattern, in four different

directions (vertical, horizontal, left diagonal anc right

diagonal) were taken ana the Fourier descriptors cf these

profiles were used as the characteristic features. The

system was trained using an adaptive learning algorithm so

that characteristics from different handwriting samples

could be absorbed. A linear difference polynomial was used

as the discrimination function. With the 128 samples of a

16 character set, a 64'/. recognition rate was obtained before

learning which improved to 9c/. after learning. The diagonal

profiles represented the information in the curvilirear

components of the characters and in combination with

vertical and horizontal profiles, this method yielced very

encouraging results.

TElUGU CHAkACTtN KfcCCGMTIcN P a ) e 7 0

In the secona method, a nethod of obtaining

cross-section information in the rows and columns of the

pattern was used for feature extraction. The same linear

differece polynomial was used. A very low recognition rate

of 2C"/. was obtained. Improvements were trieo by obtaining

condensed cross-sections and doing noise removal etc., but

with little success. Finally, Fourier descriptors of the

cross-sections yielded excellent results of 82'/. recognition

rate before learning and 9?Z after learning. A comoination

of tne Fourier aescriptors of column and row cross-sections

provided unique r epr e sen ta i or of the characters under study.

o.2 FURTHER STUDY

This study used only 16 from a set of more than 2000

Telugu characters. An obvious extension to this stucy would

be to expand the test character set to induce more complex

characters and characters of different lengths.

This stuay yielded excellent results with Fourier

descriptors of column and row cross-sections. This can be

easily extended to cross-sections of left and right

diagonals and should yield promising results since the

diagonals would represent features of curvilinear

components.

In this study, the use of conoensed cross-sections did

not yield encouraging results. A more promising method

using symboliccross-

sec t i ons , usee by Sirorncney et al ISir

TELUGU ChAACTE. KECCGMTICN Page 71

78], is worth examining. In this method, the normalized

lengths of the various segments are maintained.

Another interesting and useful applicticn would be to

develop a Telugu script reader. This would require, in

aocition to recognizing individual characters, a way of

identifying a word, with contextual information. This might

require a knowledge base of rules to compose words and other

structural information.

Telugu characters contain many curvilinear components

anc are composed of a basic set of primitive symbols using a

regular structure. These characteristics suggest that

syntactic methods are more suitable for Telugu character

recognition [Paj 77J. As discussed above, these methods

were considered out rejected for this thesis. Employing

syntactic methods [Sin 83] would te another interesting

project.

As another project, an elcetronic fascimile reacing

system can be developed. This would involve compression of

the original image and its later regeneration. Fourier

descriptors of projection profiles could give a compact

representation of the image. For example, a 64Xt4 byte

image can be reduced to just 32 bytes. Ccnaensed

cress-sections can provide even more compact representation

(< 32 bytes) and are much faster to generate (no need to do

the Fourier analysis). It might be possible to recreate the

original image from these compact representations.

TELUGU CHAkACTEl REClGMTIlN P j g e 7 _

The objective of the thesis was achieved oy comparing

various methods of feature extraction and discrimination.

The condensed-cross section method with no acaptive learning

scheme did not give encouraging results. However, excellent

results were obtained by using magnitude spectra of

projection profiles and cross-sec t i ens , with cross-sections

scoring slightly better than projections.

bldLlCGRAPHY

lAnd 681. R . H . Ance r son , Z yn t ax-d i r ect ed recognition of

hand-printed two-dimensional mc t hemat i c s ,

Systems for Experienced Applied _.. . ..

I-.Keller and J. Seinfelds (Accademic Press, New York

'- Interactive

Mathematics, ec. by

1S63) .

IBar 72].

Computers,

D.I.Barnea, h.F.Si Iverman IEcE Transactions

C-21, 179, 1972.

on

[ben .34]. M . 8en-tias sa t , L . Za i nd enc e r g 'Contextual template

matching; a distance measure for patterns with

heirarchical ly dependent features', IEEE Transactions on

Pattern Analysis and fa chine Intelligence, vol Pa'^1-6, f.C.2,

pp2 01-11, March 1 9 a 4 .

[ 6 h a 7 2]. 3.K.iihargava,K.i.Fu, 'Application of Tree system

approach to classification of Buoble Chamber photographs',

Technical report, Z.-FE /2-3G, School of Electr. c n g . ,

Purdue University, <.Lafayette, Indiana (isov 1972).

Ibil 6 01. J. A. \\ iles, 'Adaptive recognition and synthesis of

transcribec

University of

Jazz Solos

Kansas, 198 0.

by computer', M.S. Thesis,

I b r i 7b]. c.U.Brigham,

Prenti ce-Ha I I , 1976.

"The Fast Fourier Transform",

[Cha 731. S.K.Chang, IEEE Trans.

Cybernetics SMC-3, pp.257, 1973.

Systems , Man ana

[Cha d3].

G . S i r omoney ,

Tarn i I

Confer ence

pp. 1220-4.

M.Chandrasekharan and rt.Chandraseknaran,

'Context dependent recognizer for hancprinted

ers*, Proceedings of 1983 IEEE International

on Systems, Man anc Cybernetics, Vol 2,

(Cha 85]. r . Chandra sek her an , r . Ch an dr a se k h ar an ,

G.Siromoney, 'Computer reccgnition of Tamil, "lalayalam ana

Oevanagari characters', Journal Instruments Electronics and

Tele com. Engg., (India), Vol 30, Ml. t> , pp .150- 154t , Jov

1984.

[ C h i e 751. Y.P.Chien anc K.S.Fu, 'A flexible template

matching technique for boundary detection', Technical report

TR-EE 75-4, school of Electrical Engg., Purcue University,

west Lafayette, Indiana 4 790/, April 19/5.

(Cho t> 8 ] . N.Chomsky, M.Halle, Thi

(Harper ana kow, New Y o r k 1968).

Scund Patterns of tnglish

TELUGU Ch A ^ A C T E -rttCuMTIcN P a u e 7 4

[Chi 80]. P . Ch i nnuswamy and S.G. Kr i shnarr oo rh ty ,

"Recognition of handprinted Tamil characters', Pattern

Recognition, Voll2, ppl4l-152, 193C.

[Cia-1 831. Steve Ciarcia, 'iuild the licro C-Cam

Solid-state Video Camera, Part L; The IS32 Optic Ram and

Micro D-Cam Hardware', Byte, Sep. 19c3, pp. 67-86.

[Cia-2 83J. Steve Ciarcia, 'Build the Micro D-Cam

Solid-state Video Camera, Part 2 ', Computer Interfaces and

Conrol Software*, Eyte, Oct. 1 9 b 3 , pp.2o9-281.

[Dat 74]. A.K.Datta, 'An experimental procedure for

handwritten characater recognition, IEEE T r an s . Compu te r s ,

23, pp. 536-545, 1974.

[ L a t 0 5 J . A.K.Oatta, 'A generalized formal approach for

description and analysis of major Indian scripts', Journal

Instruments Electronics and Tele com. Engg., (India), Vol

30, NO. 6, pp. 155-161, Nov 1984.

IDeu 673. E. S.Gputsch, 'Computer Simulation of a

cnaracter

recognition machine, 1 and 2, Post Office

tlectrical Lng. Journal, 60, pu. 39-44,10 4- 1C9, 19 o 7.

[Did 76]. E.Oiday and J.C. Simon, 'Clustering Analysis',

pp.47- 9 4, Digital pattern recognition ed.K.S.Fu, 1976

[Cud 77]. S.A.Oucani, K.J.Breecing and R.B.McGhee,

'Aircraft Identification oy Moment Invariants', IEEE

Transactions on Computers, Vol C-2o, pp. 39-46, 197 7.

(Ede ol], M.Eden, M.Halle, Proceedings of 4th London

symposium Information Theory ( 8 u t ter wo r th , LGndon 1961)

p p . 2 8 7- 2 9 9 .

(Fed 66]. J, Feder, Information anc Control 13, pp.230,

19fcd.

(Fre 61]. H. Freeman, IEEE Trans. Elec.Comp. EC-10,

pp.260, 1961.

[Fre 621. H. Freeman, Proc. national tlectr. Conf. lo,

pp.312, 1962.

[Fu 763] . K.S.Fu, 'Training in linear classifiers',

Syntactic pattern recognition, op. 7-11, ed. K.S.Fu, 19/c

[Fu fool. K.S.Fu 'Syntactic pattern recognition and

applications' Syntactic pattern reccgnition, pp. 95-135, ed.

K.S.Fu, 1976.

LFc dJJ. K.S.Fu, 'Towards unification of syntactic and

statitical pattern recogntion', Proc. on 3rd Scancinavian

Conf. on Image Analysis, July 13, pp. 13-22.

TELUGU Ci-AkCcTt^ ZCOGMTZiog e

7'

[Kan 72]. L.N.Kanal and B.Chanarasekharan, *Un linguistic,

statistical and mixed models for pattern recognition',

Frontiers of Pattern Recognition (S.Watanabe, Ed.),

pp. 163-192, Accademic Press, New York, 1972.

[Kno 67]. P.J.knoke and P.J.Wiley, Proc. IEEE Comp.

Conf., 1967, p. 142.

[Li 83]. h.H.Li and S.C.Chang, 'Projection profiles and

Fourier Transform for Chinese character recognition',

International Journal Electronics, Vol 54, Nc.2, pp. 299-309,

19.3.

ILuc 631. D.Lucas, 'Moments technique in picture analysis',

Proceedings of IEEE conference on Computer Vision and

Pattern recognition, pp. 178-67, 19 8 3.

Mar 84i. S.S.varwah, S.K.Mullick, R.M.K.Sinha,

Recognition of Uevanagar i characters using a heirarchical

binary decision tree classifier', Proceedings of 1984 IEEE

Conference on Systems, *i a n and Cyoerretics, pp. 414-420.

[ "1 ca 75]. B.Moayer and K.b.Fu,'A syntactic approach to

fingerprint pattern recognition', Pattern a cognition. Vol

/, 1975, pp. 1-23.

[Nag 721 . R.N.Nagel, A.Rosenfield, Proceedings of IEEE, t> 0 ,

p. 242, 1972.

[Pav c 8 J . T.Pavlidis, Pattern Recognition, 1, pp.165, 1 9 6 H .

(Pav-1 721. T.Pavlidis, Pattern Reccgnition, 4, pp.5, 1972.

[Pav-2 72). T.Pavlidis, 'Structural pattern recognition',

Frontiers of Pattern recognition, ed. S.Watan3De, Accademic

Press, New York, 1972.

[Per 77]. Eric Persoon, and King-sun Fu, 'Shape

descr imi nat ion using Fourierdescriptors'

IEEE transactions

on systems,man and cybernetics, Vol S^C-7, No. 3, larch '77,

17C-179.

[Pfa 71]. J.L.Ufaltz ana A.kosenfelc, Proc. First Intern.

Joint Conf. Artificial Intelligence, Washington D.C., liy

1966, pp. 609-619.

( k a j 77]. S.N.S.kajasenaran on a e.L.Deeksnatclu,

"Recognition of printed telugu characters', Co.iputer

Graphics and Image Processing 6, pp335-360 (19/7).

tkay 85]. A.K.Ray and . Cha 1 1 er j e e , 'Design of a nearest

neighbourhood classifier system for Bengal. character

recognition'. Journal Instruments Electronics and Tele com.

EnSa., (India), Vol 30, NO. 6, pp.22c-9, Nov .vo4.

TELUGU CHAkAlTLk ZCZMTIZ P a ; e 7 o

[Ros 711. A. Ro senf i e I d , J. P. Strong, 'A grammar for maps',

Software Engineering, vol 2, ed. J.T.Tou, Accademic Press,

New York, 1971.

[kos 76]. A.Rosenfielc and J.S.Weszka, 'Picture

Recognition', pp.l35-luo, Digital pattern recognition,

ed.K.S.Fu, 1976.

[Rue 79], F.R.ftuckdeshel "Frequency analysis of data using

a microcomputer", BYTE, Dec '79, pp.lC-35.

[Sha 681. A.C.Shaw, Rep t . SLAC- 84 , Stanford Linear

Accelerator Center, Stanford, California, March 1968.

[Sha 70]. A.C.Shaw, Information and Control 14, p. 9, 1969.

[Sha 70]. A.C.Shaw, Journal AcM, 17, p. 453, 197C.

[Sin 7*]. * .*. K . S i nha , H.N. MahaCala, 'Machine

recognition of Oevanagari script', IEfcE Transactions on

Systems, Man and Cybernetics, Vol. SMC- 9 , No. 3, Aug 1979.

LSin 83]. R . -i . k . Si nha,'PLAivlG -

a picture language schema

for a class of pictures', Pattern recognition Vol 16, No. 4,

PP37 3-333, 198 3.

(Sin _ 4 ] P.M..K Sinha, 'A knowledge based script reaaer',

IEEE transactions 19 84, pp763-7b5.

(Siri7

6 1 . G.Siromoney, K.Chandrasekharan and

r .CnandraseKharari) 'Computer recogrition of pirnted tamil

characters', Pattern Recognition, vol 10,243-247.

[Son 651. Oer-tsyi Song Chuanc, 'Adaptive pattern

recognition of handwritten Chinese characters', M.S.

Thesi s, k . I .T. , 1985.

[Sta 78]. William D. Stanley, "Fast fourier transform on

your homecomputer'

BYTE, Dec '78, pp. 14-25.

(Tan 84]. G.Y.Tang, P . S . T zeng , C . C . F su , 'A microcomputer

system to recognize hanawritten numerals using a

syntactic-statistic approach', Seventh Intenational

Conference on Pattern Recognition, pclCbl-4. Vol 2, 19 6 4.

(Tou fc21. Optical Character Recognition, Spartan tocks,

19t2

[Tcu 74J. J.T.Tou ana K.C.Gonzalez, Pattern Recognition

principles, Ad c i son-fces 1 ey , 1974.

[Uhr 63J. Leonard Uhr and Charles Vossler, 'A pattern

recognition program that evaluates and adjusts its own

operators', Computers and thought, (E.A. Feigenbaurr and

J.Feluman, Eos), McGraw-Hill, New Yor, 19fc3.

T _ L U G U C h A r. a c T ci-

-'fiCCGMTI.'v fdO c 77

Z.Vassy, 'Industrial pattern

.a syntax aide approach', Proc.

First Joint Conf. on Pattern Kecognition, Oct. 1973.

(Van 73].

recognition

T.Vamos and

experiment -

[Wid 741 . B.W i drow,'The

2, Learning Systems and Intelligent Robots, ed.

and J.T.Tou, (Plenum Press, New York, 1974).

rubber mask technique', Part 1 and

by K.S.Fu

\=\Gc A--A SAMPLE'S p-opa. |&-_.

&U O^J OkJ (X-}

&3 CkJ CKJ Ol)FlcV 4-5 i.APPLETS P^ 3&4-;

Flc3r 4-&

ex

rt*npiers p^ g-^g

/A

D:_?

O

ttGc 4-7 sAMpuSc POfa. "--8:

cu cO CXJ

cO

pf^_"~4.~i__'-

r^PL-tfe PdP-.~

<>\e^rdrrO

cD o^J

oU cU oD cxJROc 4-^

sA^Pl^s POfe\\ _V-1-___L_J

"PkSr "^-7D~Trpn,lPt^s:"p^---T3-^-r4s

CO oo

CO CajF>Gr 4~>l

_cAMpu=-- -Fop5~8*

t>AT^ <L-TCiOCTUg.ES

STO DAT

FD or \ ETRT 'CAL PPCTtCTc^ PK"^P'I_E 1 (32.>

1 CS2 1

i C56 ">

j C56)

.22.

Fr> or MOEM?-/UT*L

Fr> nr LETPT HslA-VOUftL

F t> Cf R?C^AT p-'A<D*/AL

COLUfAAj ROIvJ VALUES

i1

J

PaDi^ S2c7r0 VALUES

wtn&w OP VERTlCAt. pftdJ pfiOPUF

ROfcya-

LSTFT t>l*_r

1 PlfirHT P/AG 1 52.1

!r

i

eOloMJU puw '56

<S9J*T..

poto sum 56

_f_ efcT\*A*-.

RNdW"

bi*S-

COLOMW RUN

1 ROIO (StUfJ

48-5

S*MPL.e=

-DAT

5e5u_njt

4--12-

3

J

au

\HiN*ii,inorsoortN(n*in

\HN(n*mfli"a(jiOH(<ifli'ino

J

2-

Q

U

4

tnotflr-rttafn<*oivorcMinokt0rt f*oocooflr-flr>*3\vo;NaCToort

t~n^csir-r*r-r-vo^'Ooo<sfio_>i cBifimrwc>oi,'i,i,mortminc

rtOt-OaicmaioiBMnvMnoi cwoiMncMmicanmricti?. Mcca

<-)rni/)i/-sooo4><4,lnj)^><4<<4t^iYisii rtr<ifl.n<yk<J>fl,**o-oflfl'*fl *>*

_

ol

0

0

+-

at-i

Ul

__

QC

l_>

a

cn

_u

cI-l

i-l

_-

'J

C

ui_>

_c

a:

UJ

u

cc

<*

X

u

3

__

LU

H

IH

X

L3

I-l

CC

?a

4

__

O

a

iHtu

Ul

J

u_ .j

rt crt

6- 2

10 O

M N

OC rt

C_! CC

.- O

u

>-.

CC

-X

X

u

_w i-3

O -t

u

H rH

x a.

ts _J

m >cu

s

\

HMrt-)inNOaOO'H!jiOUl^,')^'

mc. i/ikOicfir-mocaivocooomriir-

y)ino>'>'*->o_r(_f<i^i*i.iijnrt

OMMtc.iiJr<<\om(Mr-(ra)(Mi>.m

rt*Nin'ilr10i,)iJflNClNi'

"T"5,ifl>_C;0ll<l(m_'J,O(Mh\C^

* f* fl * ro ro ro * .-as cm * ..n ..

co-Ortflflrtrtr-fl -i -cr u> fl** -4*

-->

oti<npirtttirit\iN>c^mtir^i~

inMooooinmis^ONavaiooiw\Ovcroroo\o*rACMrocM.-i*Tfor-*'

-< rt iN "O * ll rt (N !N rt rt CM H .-S -<

^NOimr-^i^oomamooi^coa) in o o n d w in >n o i fl ,n /i m n - ,

r^moivooooMOinmo^ciiiAiniCrtici^inrtmiJnxinoioi/iaitOrtffN

cn rtoic.(M/>B)mmminrtr*rtrt

oofl"3 ot-*oi'iforiiiin'af<>/>

"3* 00 rtOCNOvO^lNNlMnr'B.O-Tint,irc>*vor*ccMvc-rncr<j,oocr^

mocl<'ifftCODff<OCCiCOM<imi<1

0\flCOC7-jOr-<l,Ortrt:j."Mrt>Ortfl

rtDOJOOiniNOlN^Brt * fM -!

.T

ooNMi(v>Dnrrnp*rtO'cca

voa>rt<Nvor-r>oov_>*oinrfOrtCMCfi M>ovooioii'ir. wvcoovot^oi^rt

ni/)1Oa>ini<l3^!Nrt3>rtrtrt'n3l fl'BOV1,rtOil'f^3V.NOrt'3>0

rtrtrtrtrtrt .rt rt rt rt CM rt rt rt rt rt rt rt ^ rt rt

31 CD (Ji i) OO) O^B^Ofl^'TOJJ

00 \OvOCTi'^l,m*t~QDCM_5(\*jDr'<*1

rt fl vO rt ST'* CO ~> rr. -t ^ ,-t & rr. S) *T

o*mooinOrt(_cott-HaitN<i'(_m

ijno>oom<p*-oroCT^vojD^

>NNrN,O'-l00^>O-4ll/lO9Mnrtr r*

(Min^tOHr'itoinin<a'rrtOMMain

j.i/iH(Na)ijiiOiija)fnrtOrtii(j>

3

__

I_c

O

_C

JO

II

o

a: an

to

*-

3 O

r in <j*

ro <t* ^*

.rtrt7t>*.oaV'NJrt(JlO rtvOrtar^'?r*^oor.rt.-o

iflOOiCJKl'nMtl'lHrviii

co cm cm

0* rt

* * oo

00 VO CM

co*^i,t-*^-,^,csrr-'^'CN vccMsDvc^ra,CMcso\CM^ii/)rtTCMo-

m m fl i in * o o oi ifl o * rt i* o rt in-11 a o> -^ r> r> <h .o .n a

CMCNCMrtCMCNCMrtrtrtrtCMU! rlrt(N(NCSrt^(\Hrtrt-i

S-

j><,Ninj>inov._*ofl>_3(*o.Ni^ai*M'nx)j>rtOoo>-

r^uivO^'0*<c*o*<'(S2Cinioa>CvCf>ifriir, ccNvccsnir. c

ll

CC fl

U cn

H rt

U=.

_s

3CC

colfiO(^lfl*^rt^^ln^Olll(_o(^<^ol^^lJlNO<^'

j_i-cNrt-in^(>incNij|*oaM>-irtCNmNoi''j5<r'N

tu -CVf-i

i-l x*

<3-o^voomoa

in o ro o fl <- -t :_

flr~c~sOflr~i~\_ji-3ir. fl^flfc-

c_

_t

M(*i*infli^j.<J>OrtN-#oiia;v(Nm'in*^uBOOrtN(si'in>Oij

rlrtrt rtrlrt rt 4 \ ^^<^i^l^l<^_]

X ) --

U (.

492 492 ^92 492 492 492 492 492 492 492 492 492 492 492 492 492

TYPICAL VALUES AF'iER LEARNING phase - CULUMu fcU.<S

0 0 0 0 0 0 0 . 0 0 Q 0 0 u 0 0 0

Q 0 0 0 0 0 0 0 u 0 0 u 0 0 u 0

0 0 0 0 u 0 0 0 0 0 0 0 0 u 0 0

0 0 0 0 0 0 0 0 0 0 c c 0 0 0 Cl

0 0 0 0 0 0 0 0 c 0 0 0 0 c u U

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 Q 0 0 0 0 0

0 0 0 0 Q 0 0 0 0 0 0 0 0 0 0 u

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 Q 0 0 1 2 0

0 1 1 1 0 0 0 0 0 0 1 0 0 1 2 1

0 1 1 1 0 0 0 0 0 u 1 1 0 2 3 1

1 1 3 1 0 0 0 0 0 1 1 0 2 3 2

1 i Zi 2 0 0 0 0 1 1 1 3 3 3

1 2 3 2 1 0 0 i 2 2 3 3 3 3

1 2 3 2 1 0 2 1 3 3 3 3

2 3 3 2 2 0 2 2 3 3 3 3 i

2 3 3 2 2 0 2 1 2 3 3 3 _* 3

2 3 3 2 2 0 2 2 2 2 2 3 2 J

3 3 3 _: 3 Q 2 2 2 2 2 3 1 2

3 3 3 1 J 2 2 2 2 2 _> 1 2

3 3 i 1 3 2 2 2 2 2 2 1 2

3 3 3 1 2 2 2 2 2 2 2 1 2

3 2 3 i 3 2 2 2 2 2 2 _2

2 2 3 1 3 2 2 2 2 2 2 1 2

2 2 3 1 3 2 2 2 2 2 2 1 2

2 2 3 1 3 2 2 2 2 2 2 1

2 2 3 1 3 2 2 2 2 2 2 1 2

2 2 3 1 3 2 2 2 2 2 2 1 2

2 2 2 1 3 2 2 3 3 2 2 2 2

2 2 2 1 3 3 2 3 3 2 2 2 2

2 2 2 2 4 4 4 2 3 2 2 2 3 2

2 3 2 2 4 3 4 2 3 2 2 3

2 3 2 2 3 2 3 4 2 3 2 ._ _. 3 2

2 3 2 2 2 2 3 4 2 3 2 2 2 3 3

2 3 2 2 1 1 2 4 2 2 2 2 2 3 2

2 3 2 2 1 0 2 4 2 2 2 2 _. 3 2

2 2 2 1 0 0 1 3 2 2 2 _. 3 2

2 2 2 1 0 0 1 3 1 2 2 2 _. 3 1

2 3 2 1 0 0 1 3 1 2 1 2 2 3 1r\

1 3 0 0 0 1 2 1 2 1 "1 2 3 0

1 3 C 0 0 0 1 0 1 0 1 2 3 0

1 2 0 0 0 0 1 0 1 0 1 2 0

0 2 0 c 0 0 0 0 1 0 1 2 2 0

0 1 C 0 0 0 0 0 0 0 0 1 0

0 0 1 0 0 0 G 0 0 0 Cl C 1 i u

0 0 0 0 0 0 C 0 0 0 0 0 1 1 0

0 0 0 0 0 0 0 0 0 u 0 u 0 1 d 0

0 0 0 0 0 c 0 0 0 0 c 0 0 0 0 u

0 c 0 0 0 0 0 0 0 0 0 0 0 0 u 0

0 u Q 0 0 0 0 0 0 o 0 0 0 0 0 0

0 0 0 0 0 0 0 Cl 0 0 c 0 u 0 u 0

0 0 0 0 0 0 c 0 0 0 0 0 0 0 0 u

0 0 0 0 0 0 c 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 c 0 0 0 0 0 0 0 c 0 0 0 u u

WEIGH1'S Oi? CON DfcNSED ROW fcUi\S ft.Tt_R LEARNING pha_>e

TELUGU CHARACTER RECOGNITION - ADAPTIVE LEARNING PROGRAM RtV 7

ENTER NUMBER OF SAMPLES FOR LEARNING:

NUMBER OF SAMPLES =

WEIGHTS Of CONDENSED COLUMN RUNS AFTER LEARnTnG PHASE

492 492 492 492 492 492 492 492 .92 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 49 2 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 4^2

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

432 390 404 25b 48b 486 48b 446 432 460 418 432 446 384 444 416

30b 402 416 430 486 486 48b 334 320 390 236 306 30b 402 416 4lo

268 392 406 434 492 492 492 268 266 212 392 406 266 366 41b 392

172 258 424 362 334 486 404 370 152 222 314 342 20b 344 424 372

324 310 436 412 156 394 394 43o 58 422 324 324 324 43b 43b 43b

224 308 456 448 214 348 41b 364 303 44b 308 456 224 45b 45b 45b

92 246 440 456 194 400 432 38o 386 162 316 440 31b 440 44U 440

280 456 456 476 204 204 464 28 0 378 476 456 45o, 28C 45b 456 45b

240 464 464 464 240 330 254 240 270 464 464 464 464 464 394 464

178 456 45b 472 276 288 268 354 472 472 344 276 45b 45b 374 45b

472 472 472 484 472 484 282 386 484 484 386 288 472 472 474 386

476 476 476 484 476 460 302 484 484 484 380 28b 476 476 446 406

466 468 468 376 46b 376 176 472 472 472 332 <-0_. 46b 4 02 390 402

456 456 456 488 308 362 166 476 476 420 364 420 45b 476 37b 4?0

452 368 452 492 452 380 156 480 480 424 312 424 452 48O 492 424

356 382 452 492 452 492 156 438 480 438 382 424 452 48G 492 424

426 468 460 366 460 366 156 426 42o 426 384 46b 330 466 352 46b

304 444 448 352 44b 184 416 444 444 402 416 472 430 472 336 472

416 416 436 180 43b 180 416 416 444 38b 416 472 434 472 346 472

324 374 444 194 444 472 444 374 430 31b 374 430 44b 472 16b 47 2

424 368 424 212 452 48O 452 368 368 452 452 424 352 48O 424 480

414 274 414 240 456 484 456 456 414 456 45b 414 492 484 344 484

404 278 446 48b 484 486 484 484 446 456 334 44o 492 486 45b 460

380 40b 436 492 476 492 408 476 380 464 436 492 492 492 464 43b

352 454 492 492 440 492 440 472 352 468 422 492 492 492 46b 4?2

334 454 488 488 270 488 454 476 334 468 404 486 492 488 46b 46b

46b 472 468 468 292 250 342 48b 412 370 4l2 46b 362 46b 472 34b

466 468 412 468 292 348 296 486 412 384 466 466 23b 466 46b 342

392 354 344 252 200 300 270 456 376 392 448 446 266 448 464 26b

464 422 342 346 32b 300 250 472 422 422 464 464 234 464 472 3 52

468 484 382 468 442 442 192 484 412 468 384 46b 202 466 484 326

464 428 484 402 472 444 168 298 356 464 372 204 464 464 484 26o

456 408 476 472 472 402 194 282 304 414 272 28 2 400 456 476 252

372 398 460 476 476 476 194 300 290 390 262 300 370 460 460 40b

384 384 466 472 472 472 272 374 306 420 252 420 420 466 46b 360

448 480 434 480 480 480 480 410 318 332 410 294 4IO 472 420 48U

308 480 44b 480 460 480 480 480 318 410 410 28*. 410 436 462 480

276 480 410 480 480 480 480 480 296 438 256 424 424 456 450 480

294 484 428 484 484 484 484 484 4 42 484 484 442 484 476 284 484

470 484 470 484 484 484 484 484 470 484 484 470 484 262 27b 484

492 492 492 492 492 492 492 492 492 492 *9'2 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

492

4<>2

492

492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

492 492 492 492 492 492 492 492 492 492 492 492 492 492 492 492

-FDSor PRqj EcTIO/vJ

CROSS.-SECT \OS-S.

5.2.

PPOFIL-Ts9 FOR. 16 STT> CttAJ5r^CT)E7$S

'/&

t* ^, *rt - |l |S(V w * IT. O *>

*tfifr r-c*> tri ~~

Ar-n < # j m or*r<o

r- n cr it cm m c

01 w* m ***

vtOISiB

O ft V)

tr-m^r. a -i (sr

tc i CH

m m m o * * I

i <4 c r win

in ^ u . m o

itnrs

a_ <Mn m ira ii

is. in - r> cd r

- tn *- m

in* m ith

tn m r-

*n r- m 4 ^ c-*

** o ft mo v>m *t -*

Il-Ki/I *-* * A O nt *rt a> r-CB tOA

f o>Mr ocdcm ** ** r* ** m *~

>f*tnv) aa r * csm*i* *>

or--/_> ji ji .-> in oiMvf o m - onc *w cr if> c cc m ir. o "* r- e r*

* _

ot * * ir maa < m * m** c- c*

oi/ir r*inflr* rNo^tr <n -i> o at

B WlfllT m *C0>

CD *-C IT, O W<*l

r- * r* w in r> a. n < mm a, r in

9 o^ ifl nn mr* a am a jiuiocd m

moiO^v a tfiff< wibm ir oim

_ ,

BflDMo r* * <* m m<MM r* !* n o w

mm m m

c* ** v m c I1

m m m w

jf> * * -o ->

m in * _ ,

no>n a. if. r *j>

*.r- ** * *-* o in *

r_r *

>C * m

r*

P- tfl o o* cn *i in

o <* *

AOiON

o * ff>_n

no-

o **

IN C

*r r

3 4> a

<r m cc o%

m \ rv* .n

CN .1

tn or * .-*

Bin -._ii^^ rt ** m ** k

I- oiiniN * -^ *n * r*

av r- e m m tcr-*c _

IWw* u.

(\amc

C OBjB.

. c r-a >c

!

C 1*1 -" - -i. -f*v v a in i* < * in ir. o e*. c*

c\ w vf o o r- * r- c o. tf c^>

in tr r- i

* r c

U" * * 9< *i* a- tr. w in * .

*t in m

'Hire :i n o n X <e HMhN<

' i* *f m * ;

^ * w IT I

1/I"1

(\ -* ,

id***.-* r 8

* >~i

,OC 4 M

3 .

a f * cn

ft at in r- c

__

HC'Xl CN I

IN tc tfl *- ^ IN U. 0>

CN * IN C n * in

ir. m h (

* NJ '.

(N U*-

-i - -i

< (-. c a r- ^. r-

it 1*1D"

a*

*_-C'--C'**C CNCV*CCIN'i*1 r- * *n

C T W. U

-'

n C

' *- ft tv tf. ,

.U J . .

h, ">a v i

j t .,t-i-

ri. 7 l fvj ^-i

-0

u~- a r < u^^cf^ocZ1'

I . . A > - J

4 it ** rv rt (se ii

t r .if*>

. - C T or (N

u t ir

:"_ / |Vi<

a. <n o tr

_jr t >* r

iro"r

q C\ w, rr ,;

F "^ IJ

CV C

- -< <

O IN <J

a r* * r*. n

o.-

: ^ - - q .1

i>

, ^ -I T - -* J J

a rs. 4 -< o% r ^ - i in^ir

t-i -> 3 - >> , j-. / j% .- _: ? -* n

ii a * w Pi _: M o. ir

_-!_ - * -1 n vO MJ >~i

O.J"1 o ** U- UL C

3 -^ -rt) II J

tr o .: <N IT

j * J

C 3 > --*- a. __

Jl O 1 jl 1 II 71 < >* cn j r-

|3.If- CO O* CV

,i _: - -1 J

i-i- M*>r j *% a/

_ >-. rm

X. * .3

a 3 :j

u. u

.NI aac uu.

t> tn C C7> *n c c <r - ta*

<

^ r* IT CL *r IN m O. IN ^ i

t* **

If SEJ* r *rtj 3 * 4 O /> w O Ji

-J o 31

^-

7 f - rN r- ti ar X n a

1P-|

o> c 4. 10 _T c z. M u. N.

"iJ "t i

-J Ci a.

.4 r. ...

u. a

NJ 71 3 a * >i * 0 O 7_> N

f. tr. %c R. c '" P- cr "' s*. m CN _r C r.

r* (*> ^,

IT. m C

'rt -4

CN l-

**_,4

cl_ u __.

-*-1 1

Ik- L

n o J * _> D n f p-

a a

i^s-

_. CV c cn r. VI CV |N r> ... ,4 -j t

T m -J ^ t

a '1 . <

j L c i-i*^ 4 A * 3 1 N) n t

r. ^ i C ir a f) u. f*l -

rt)

? O* c o - V in m o CT* INr* <,*

^4 4 n,_

NI ^* *

i'->

f t! -H JH a .N j i 1 c T

fr Vcr- r. If. - *f IX a ^

*

'I>wr- CN 1. c if t <J 1/. _r "*

1* *l

tr ifl y 1c r

io o t*

n -1 i -1 o * 3 N r a *

u> o- r* a. r*> r*i l~ * i-. u *- _1

-rt

>. CN ** i m r- IN -T a c i. r CN a. .7 4rt) --

v.

d

-J1

ir.

>

V

j

i_ ia ;t > r "3 > 1 r-"

r r- f a. a a a*<V n v 6 >* a a

-I -JM

ur c 3 c tf vi If !> --1

rt) X ** t T .-, jj

JZ 5 __.

or* c*

cr*n a ^ tn cn vno

ion a <- c

t> r-m *- m - N7 IN CV CN

,1.4*3

-ft CN] *> zo

tfl 0* *" fffllfX" m*lrt (,.

^O^N *t m (* tf JBHI^ 4 -ICS| D

t f> mm IT ^ ff n 9 * (B *6 W Oi C "

Oh*o r- c ** en cn o o cn cn *- **

- cn r- m

m r. m in

fatiit*

cgo:

oos . o n . ji ") -- o> O-<0>oomc w*o m cn in -< >m m

<*

Min m V CN * ft w <G CN *-t * mc m *-. * ,

nfl. f. *r*( ON- CW^w -. C-

I *4 rt A S > O> *4 a m c- c m in

> o> m cn q<c 4 rv in cn cr ir. r - r v in^ *i r-*

oc>i v ^ m * ooMfi HMn w m in m cn acr-cvo

ac cc fi cn i-

r* m * in <c a

OD m IN 1*1 i an m v w*w <c ai m cn *MC O v.

C\ * *- *rt CN CV Cr* o i" a r-wc-p*

ir. cn tn A m c

r- IT a < is f. CN . I tr- a-

<>

m r-

- z> * # -0 i f

*r- in r- b nt

1 A rt -* rt rt H fCf- (NO* IF CN U.CT

. J .

lN *-> rt rt cr O*

'.n n * -n a. rti

ir c *- cc i

O O T *

* CN n n c

-rt rt

^

IT U r-

rt rt rt) rt

- m V IN cX

'n r*

JIN

m CN CN c ci

mmftr-(r;

.%

r*cr** r- ir. rfi c^ ti

f r|VC CN cn If1 IN

1 *-. f CN

*n > *> n

CN tri C m

+ n cno

r* * * j- * 3* .n 3 "NJ

CN * vt rt rt tr 4 _T

H

* IN C O Nt w tn f O IT

4 4**

c

c:

n

-

NJ

N -rt) rt)1* n 0 -N J J

m or o m ir- - MT CC

,rt -

f*l (0"

LU* ^

IN

3 rt)

CN

II J F

"

t 1 li

0

t> ftN

>

2

"

iN

a

vi

r- ii a

*lIN

3

a

*

fl

J

4

,r. r. r

uN a NJ Ca uu.

<ft rt) n rt z>

tn o IN Cl. C*l

r* IN ^ X rt

I -J

V.

I*

o T NJ r-

r- r- m a. <r

j

ir. c _- ?i Vrt

-rt **

5

y

t If / t 4 C i/j

**i r i o. c

r- "3 N

(X O If w -

7D

icopihui ci a. ^

J -.rt) rt-4

oi *- r-c*irt cn/*iP-r-

.rt rN -*

-i -* .n _; n .- vp r* -* -H -

a <^ tr cr iro- in ? cn

if.*iL,fr^'**'"

** a * Q -+ -+\ a J* t} O 0 "N 1 i|r a c, J

h<i o<_vfVU mem *->--

r* C C CN

^ -I

r> <i ^ iA r> a r-cir^ in m CN w- <rrc

(/: -J t C.

_J-i < a.

JUG

i d hjiji 3 r a -4 a .* n o 7 -; o o n >

l, r> r> Mv. || O r n li t * -U C m vf -t "T

^^.J ort

li; CN CT lf> Mf ^M- a. >C CN t - I* r -. o

-* J

c o b.

,; i J ->

O-n-Nir-^r*a 'j -+ ~+ > od "*! T* *N r- ._: n **>_

[' o r m< a u. >x eve _r a Cm* m a, a *i r-

Ittf'C^^^C--' 1 * t <-i ^ _ m ir. r c

1 -<4 ^* -J vi -t -rt

o *-. *

u

jio-3^ t n ij- no > o o N - N ^ i <i

w-cor-^t-^^^ir r*\ ar- r* t~, m a r-

J

t,--mp'^c t--it(t __. o tr. c< - rv ri tr. -

r- rv- ^

N) -rt *4 f

r- fv ~- tr-

no m

A rt *n r* ^. j*

vl oo- m iro

vn i. V 4 K IT .

-rt JlN

rt)

_. b"_*

"N"

-n 3 r- r.

*-t IN C* fN

"

i .

-rt rr, V f^ *"

~ ~* - _ N

C

I.

* -- T J ^ Tt-

vl e a* x c* r- c

-)>

_r bC- C* o m l. vjr

-> rt) - t: *

C U v> i/ * Ol

t c a- "' I ^ "

CO

'N

o-*i-i'*iia.""J"*

mocux*-.-!!'-_-*"" ** " u<

-!_) ^'O "N -I

*_>*CN(. J V fi < 1* a

.-H ^ -4 rt .)

>rt rt ll vj

Ji O _"

d .: * "

- - rt r* n -n r- 'N -* -J

c * <o a tf) u (- j*"

'*.J

^^-rt3--^'-*-*i

*tncNK--miN^-a o.or *i ^ i

,i:-^ r

5 O

"

=v j

~

- i: J

O <>l r -. 7 'I r\ 4 3 *i *

r- ^ J P L ? *. m

in a. m

r.

j

rtl

* C-

J

1/

t)

vU

rt|

rr, _> r-

r. o a

i t 1H* r ._ o fl* -rt rf

J

if.a- **

a

r f CN

c

J

- CN -

* o t- u. r-

a - o "i

n m c c **

O O._. C (N)

vi L. CN C

- -rt - 3 ^ -^ l|

- M

U fN CN [1

o v t- r.

3 "N "J - I

F<^5-2-

m mm ov

cn mm r^ rt cz

<* INN CM BCCOir in m m r- r* * cn ^*mo cn m

v * -O

m cn i

l rt CN rt

r-t-. on A3 m ** o r*

off **r* * *

r-

,a

mo M"i tnm-i

o t in a* * j

r> <

--*

m rt cn uc vmrnm <c^**e

._, _. _ ._ 3--0-.0** rfl .A *rt)

a a r> ** r c <ec- m ire <r_>

p- in cn

ic - - *-i m cn *r p- ^ * m "

ic idaob cn mtMnir

vl cn m -*

fftra * * m o o .o * o

^.p. CT-rtp"flO* f* fl l*>

, ,_

-i in r-tn-^ft ^mn"- rt art rt) (N* "CC ^ ^* m f m o

m mr"-rt r- * * * m

<r CM

O B fl* M r*f "-I*r- a r crcert in cn o

|MOMr* m m m rtj> ** m , p-rtrt ao m

r- - o

r pi a* fa-rt)m u>i/-3*4 rt - 51 j -h

C>(Vrt{| w*-rt r-ortiC * na

v4<

r o* -i o a> cr- tn tn ft ^^m -*

*> o* i- o> i>

- <r c o- *

rt |- <

rt-jri-t

a c c a

- m r o f* r p" ooa*-- # o _

m * tr m i- - * tr. a r r- pm o* o*

mcv>rt m-rtcc r^^-ft CCvCcv r- *- r tr *C rt O

vt m <v rt

m li a a> t< q <r i

"CHf)rt*<-, ifcr o _r. o

r. q j> r vJ- ii _J ii '3

: rr m t, rt c * f. -> f> r* '

, <N rt rt VJ O f

rr* *

r-xocNW-r-vCOi

_> a n -9 - <

O vf f U. f *

- I-, a i

1 o -f - f /I a

ll - hrt1 ~*~ o *ir<t~r-

** J -<- u.

U -j

-J(i*r^CN*acM

^ '. *-

tf r . r-

IO vp

iTTi *- ^ j. o in ** rt

_ O -rt -O T O "N -* O

, ir.p-cNO-ON-murr--

j--

rt u- K'rtf.flCvH*P**^

, -, -. -i O ~

ucc-Cic i* * w m r- ;:P"

rt Ci CN

I? IT <N - .

H- 0 l Ji i> -*-

09>tr*JOrt9>CT<

"CrtO -C*

*) r* Ji r U / 7*

if. a. - *-i [._. r* O

. O CN IN rt c. ff*

rt o* cn - c m fi

* M ,.Nl .N

-^

O r o - ho 3 TC'O *

vn .1 * m in it p. C >f

, . L .

c p-,n t m fv ^ m tr-

a C

CN CN

C v5

*'

a

C ^ C C IT W 5 0"*Xi

t- f C h IT !i "T'C -- a C

j.. ..-j.* --,

-j* ,n> -rt r -*- -J "N

c- a.

-j/lrtJD ti'*"N-^ 3*"'

^^(scirt- (N--cin<i

e o

O***

. p. ff P" < *C flft*P*CNT

'f ii^c ie no

'l-U t m t if

1 "N O -rt] rt) frt)

ono** p* .

o j ^ r-o i < rt -n

(%,* O itf CN (N

"-r^ |NCfP

sf' C

ir. O-

--vi">-f"C*t'^0

o / f> m " -

*4 rt H- O* ~* f* j ^> m cn c

O^FFrtN-.

IMi B C

J N rt

* i> n ft _, ,.

.'t'KlMmo i

- J -

' J * '

<;***-.o c.

- C IN

vt ll "_.'--

* ft rt O il rrt _3 i

.j *_. *; * a. v* if.

-1

4

*r n '

a vX c ^ *rt ct p- r.

> -rt 3

C*l

- 11 .N

40

i'4 "*

K i <) .J

=> > n c c

X m -rt .4 j > Jv v# J* -rt ^

V-i.r- ^ rt r- a. P* a

rt J <J j ^i

_F. <*t a p* p- rt ?- r- O

3 1.ir.N X *N

J *rt _ r- r. r zi* 3Q'

-> -3 rt .rt c 'S O

U C_. Ob. b.

n - "_. ^ -- ^tHt 3

r-vTi-.i- - ^ * -r i

..-!'-

afN^i^vcmiNomri

J v. ^ NI

on CN CN

PI -1 ^< >l

Ui

I -rt Z3 -rt -rt

Ol J rt CN

-. - v -h n * n -r --rtrtnr--rt:

a n r <. i\a-r-^J-'NO *

u- m ni o. f **> nv m

u'

*-"

f

j -n rt j * o -

i 1

^r^ -.^ O Vl

t* m ,i o _. m rt u.* in

. . ,1 -vl

tr rt-|-.^c^rtf-'*',*

i. .<*: ^

J 9

-( -rt ,1 "> -.

-i c t-

.J h ^ *1 rt J* -1 '->

v.-c. ftO1 p- * '1 ^ * 'f

4jJ****J" '

v< -j. I ITO' * * - *' "' **

. u n . *

-*= 2* C 3 ^

OU. *-

t m

n m ft

c **- tr *f*

;- -. "i O -i fl ,1 ;r-a**cmirt- --.- it

-- _i .n -/j n -n -

f. a o< a o. p. -* u t---

j-n ^ -< d h -i u j

l/> J _i CN f

il t = J

c _r.-. ~ > ,.;_;

d A _> -n o t Jf rtJ

. It m " IT .;*- .1.

-'

j ;-.i_t--.j 1

m>-mc_.-rtiNr * ci - * * -

j. rrt __: u C .

Ci3^ Ms

*rttfi* o*on men

**tr- * * cr

n,rriN wmcN # cn*> '

rtfN O^mcvrt r **r cn cr f * - cNmcN.

,

mo vd

cva m

r- CN CN

9iin<iim it mrt a> m cnw*m r^iNioiN v <c c o

* rt (7-CNC IN lO rt rt m cn -* men am*.* m^ftrt

asap1 in h 9t mm

, i t

.rinmtr* cncn m a~m

m a> o>> m

CN ft rt

O rt m o- s m r_> in A pi o>

r*oiic- rtmr-ai ij ac

o ww a no c vmov*. a * rt mmmm o>

-

a m m .

itritn c r* cn cn

p- *r

m m * bin p- cn t m of. a-irt p- c *

in c m rt

r * tra iflirr in cn a*

<i4n m p- * rt r-<t) * m it- - If) v v-

CN m It IN IN OACNOrt MC IN CN m rt IN If. m a-

icr- ! rt omvfi I r* rt

m itfi c i- a-.

o *

o rt m

OlAlC ' rt o m o m * m vccmcN -* * m cn cn

in o c .-- ic ir

rf> w

f c

a- cn r-

rt m _r r

3 3 0 A OOOno vain

'

r\ < a i * pi i

-J >]

^ > n 3 Ji o p^ t < ft * v n r'._, *

b

^t (T fi m a

t-t er pn in if-*cr- * rt u ft o f

_, r*i

C - IT L'vC 't rt

'/- c -< m* i

fl.

*Ml

*'

J ** m _i

tn <r r- p-

CN C < r m

i"N -; ft -* _ n

C a a "* IN

Ol >

ub

t

a o*

5 r r o *

Ir. o. cc m c CT

m J N? * rt o

<

r>

13<N| -at N| ". n o

oc ?-* IT- IN CN ?

Lv vC ili-

f I

IC IN

Ol tNJ

rt -rt)

IN ^l

--rrtO. C . flH3 i I

'"if -a f ir ;

fi ev- C >

r.

.r- f c

f* IN |N .

c 1- - c

F J 0 0 Ti t n 3 n * x rt* '

t- c m m U p- p. O p- a . -- c a ap-tr j

u

m ro-

tr P o f^ _r ^i o< \t CN

c n -\ N < N

c 3

j (L"

rt J vO n i y F .1 r "_ -tr* r* i

- m m rt r> *; V- rt ^ cc

IT p- *r c T vC P o. t~ o. rv m c CNj-

V.

<- T >

h- 1

-.

r

X

u?-

M t c .rt'rt. It ii j F n 1 r* ". F 1 J

r-'f c r- O - tr

n- ( u O IPCjv O- v>

IN II ^ w C

a ~ o. oi

* _; u b _r-

CC*

M O -

I ir **

'j-n-v.

-I

i.

t

* J* * - If. It IX ,

-i - n r* -i

c a o. c x-*

"iO"

t r. m

O *i

(4

* r* N

O- *> m

o n j F - i*'

p- a. u. . c m

*N Jf. a. fN a p. cs j.

J m _>

i/.>i

c. ^a ft D m u c

ta cc UP va if r

Jrn

X

Q

m

ft

(X- IV

.J

M m rt rt

u a vl v VJ fl. c.

rt .^ } o IS)

IN**

-0

J .]

cF -> F 3 F

a u VJ. *_- O- a u.1-

k-i C U IT * K< t-

c t ii X JT. "

4 3

rt) -*i X 3

. rt rt rt cr-

-0 NJ "*!

- 1 - rt |

: j ^t m

r- > rt :

o*

f -vj ,

r -i 3 m >n n *

p t oi ck fi o

c r. C - ii o* vi

a. -^ r- *r .

i a ~ cn <

"3 rt) **

? 2 ^>!&

.,_}

C rtf - CN"*

C i

ttCNf *ONjTi C a. -. m it in m in

amm <r cn m cn

n- m miNCNK.CNIN1

oomm ooi)i*t o ^ a

tt cn r- \>p>cc m^a-

CN a cn o rt m rt rttmn

af rt CN Ol l rt <4V IN a, f- '

miCrtoi cr-mm - rt - . cn i* m

3- Ot > Ot Hi. **)i.-r

uran m cn mv.

m rt -

m IN IN rt m p- * CN IN vt m m iv 1 x m m * cj \ m cn 11 .

'

iPtncNCN m*rt*rti ifl PN

rt P- Plft O (Nr

<T -1 vC CN CITiTi

* m oi * a-, -f if J

* m vi. m CN CC NT C

vC fli IN C -firocNrtiNCNCN C < O rtO'*--

-

O rt CN f. Vt IN .

n _i n m m o z> ft

f ri rt a- a o- <r <r cn a ac

Nrt rt JN

rt IT rt rt tr vc *r rt

rt rt -rt rt -rt

(\ m n ffl

J. 1 O D

CV <_ O -

a -or m

r- m <r *>

a

I- * CN CN

n o "> c

ru j in m o i in cf tr o* m m :n IN iT

u - m to ff ifi

J rt

qi\h-m. . rt b. 4 m ft IN

OT-rt-rti I^OOFV) MltitjJOI'Ort rt

9 a* rt o* ov cmu*-1'* o: f a tr

a;p*rtjmu.rta,*^rtvi?ircNf-.fip'0.rt m

oi o p- rt

q m p ft ia a> m i

3-3 3 * J

f in uc*

o >

m f hpi\'

vi c a i

<- fc- no

^ _b'ft n

i. T C >f IN

o . . . .

ap .

__

j r -* N aur-rtrt-.rt'rt^r-

prrjF>(fl7r"iifiC<cit*

____;_.--

~,np'<-;wrt-iN'__*rt|vi--nr'P-rt

Crt >. o(- a c

i 'jC U vl U

- rt 3 -i

n 3

ri Cif *

< n

a

.1 N 1 - sT

fN *4 h itlD '

"T IT 6 P" **

[Vl .NI

"N o __ a ~> rt

vCP- b - Ni IT

> "Uin m w cn ^

rt4-

Cl l"--1"T IN

tr o *

x

3

.. ,- |/v

t C-4 n rt .1 rt

t- m tea-

u

i uu. - r- ** vr i

i rt -3 .vl

t C

J 0 N O J o o >* n JONT-I

. r i\ in a o u r C f<"

c <n

t if <" ? 4

. . IN

c

i ii C C V I

If IN Ol

-

m fi c U -> **- N- vC

__ c c

4 j> .- a a ^ T. r- n _> * -rt rt m -u N N

r-uirtTuifiCir-H.oo-

C f IT- ITi W C * t- M 1 W *

P' Is

., 'I N " N

'if, f; t vf

O Vf

C vi. vi

i 1 rt .

n.a-

ir

II J N

!i

h- n o

"UJ ii .

" ne if

n t: -rt- 4 _"'

XT

Ub.

P * IN vi

n c o-

N

If.

Q1 n um f m<*w n. m vr

-rt

N. "T U. O VT i

j rt

IT,

-rt r- > r*}

O: rt a r-

-U

p V I-l ^

x *n

rs

b.

C v C if . VC f

ft ._ 5

**-. a

V*

in C vr

> .- - i* 3 Ty nj

- -- u. c> .' f - :

-4

: m a a h in ^>

u .NI

- r* O -. rt .- irt^ _Tm ft 3 n. I * -. 3 n J ft o o r N : 3 o rt -

^rt -.

rt^N??<' cnc- miN^m r-^a ^

i_"#pn,*7/:.,**-. ^ ^1 S^

-/oTc'rt --<--rtrt c;^'iCta;/'^Lr'^''-ca/ff. rtrt^- rt art

^

"

_ - ^ . ..:* -# rt - -. -

*--* - Xc j. o m

" .*"*"

S -rt ft o r-

"

T ^ * rt - . n o ft -"- n

J-

. ^ - -. n.- ** ^ - ^ ^ " * '

^ c, r, c. ". r, . . . . j n vj rt ^

. ._ '.-.-.iMLn J rt J rt J **

.1 O J ft ." *

C 'f vc -r v cr- if- vf p. c

rt -* 1 rt -

o C C'

Ti 1 r? ^M ( N

r 4t f vC it vt* i_-

?rt 'rt -rt

a. UL <c a ^ Or-

Ci. -T

,_t rn VJ n s> rt

m r* Tnu

cr _n ur- p* o

p* m C J *

t n

3

D -H Ol J -1

vc n m (& o^'

jtr. *- O' fl rt I

1. o*

. a CJ O 3 N

P-

- Q. CDP* I*

-f

rj J4 ^

r- in a f

j; .**?

. 4 3t X 1-3

U b-

tf)

I .1

^ofFO is-io-. ftoNnr: n -hi

% _

^r-mo-m-r-c-vrf---rvif. -- fa -

li rt t .rt;**

^ _ ^;.-; S 5 -> ^ j r-i -^

2 u. u a

H EL o |N m m O. o r^ tf- b. p rt

-J _i J J

4 *N r- a* ii- a n c vC CN

X c rt" X * _c

*r "3 .->

X o a O

U __. __.b.

PCq S-__ <&|&

8 **

rt m <n ot*u- # _DCna)j_ir* <N)m j.

crco-rtiwrt r* to 4 o cc <r 4 cr c r a-> m_ *

m r * *- e 4 rt in NrcmcN 4 4 *

<> m rt *rt iv> in 4 O jd rp- m p- c <x m m

y"

c rt

**

n? oi * cn m m in c v < rt

m o

tea %D m

OvONP

Or i*

t -

CD HX-iO

a* r** -rt r- <o o r

mmim CN rt ac <c

r> o m cficNCN *c cn cn m

s rtrt

rt m 4

m in *- m 4 m m ev in m rt m cino <- tr 4 in

<"rt)m ^**om rt c?i s> m o^ it a> PWP CD CM If) 1*3 CC P* in 4 Cf rt O CD

trt <c m in m r~ 4 ai O #rtm CN a- m oi m m ft

V o o o

. cn m rt

B (vj n * *

IHi c*e cf o

mm votff m OD CN IP rt

| ft tf m * - po.cn r- *n m cn i m 4 rt tn m m rt

c 4 m ct

r- m ir *-

0 H1f4 3t

fit *. miNr-

Ct C* Ji <H 4 4 -* P-(0-I0>0 rt -rt m

- o* v* c rt g rt cn o9mf> r- m m

o> a. > m m 4 m m rt in m o p- 4 cn to cn m cr in m ci-in o>mcv,rt

4 P- O-

4 a. 4 m

nn as r* om m rt rt m rt r* r* ift r- cn p- ,-n 4 r o :n m

i m m ^ cc rt < oi od m in * cr 4 o

PNrrt iiw* art4ma- r~ m rt cn cn O ifi CN O U> o f 4

j < * C CD CN rt 4 m vc m p. 4 c iNp-o r*CN* vc vC m m irr i

CN art

m 0 p o *- O*IT IN f> mm

a tr

IN IT 1O" f" f " IT CN #*

f. m o. vi

U -I'd # 3 rt -n1 - rv f a tr

j >* o r* o

vTmP.mu.aiir mm> a. . .fi j. ^ n

a p-

cm f

o ft * 9 -1 QandAO-rtT

ira mirif.voa.o--.^-)4if i -.1 in c N <n

1 C 0 c vd c m' t

.

*f ilO(N a- -.

NI 71 rt 0 3 m -0 -ft ft "T J* O t * -*

4 vC c a 4 a. P* 4 I CN CN m r-l_ IN IN

-!_ J -H

4 vC m m 4 4 * if: - * O. (V t- 4 <X<

m cv m *n

."MM 1 i

-rt -a rt rt

CN CV f . (N

<r 0. 4 vt

1 C rt It P*| r ^ ^

p*io.rt oo;

< 1 3 3 O -. -, rt

.j* rt) 4 +.

-|..

.

~ C

.^ .* CN# rtC-

-ft

' ^ *

Vi2Srt"Kz^

l-l 0 O ft 1 'J -.1 i rt P T vt * -r : rt ^ rt T

- p- trt VT CN (_ v. 0 P- Cto-

4 /N rt If. a oi ir tr. O -

L) P- m f C tr 4 If 4 -^ **. _~ f.

r N -J M rt)

C1.

0. 'f. %

0. ^j t u"

"> * at rt n 0 r* U F * rt "1 X l-

4 iTP- 1- O M m O u vC JU CN -.

r- f. CN a

a rt ."-i*+

C o. P- 4 m C m 4 vt fi C vf. ff. 4

-4 n N rt X

1- B- ?- Tr~

c J

rTi P- -rt ft a -I ft _ -N n

<. 4 4 f> O* vl vC n. O 4 fO* P* CN * _p ? *

crt 4 -vl

f- f f

c- ^ 0 rt rt

I- r o. f r-

tb'

vt -i

rt "3 -4

C

J ff- -

I

UC"

I

-rt 'rt

C"-,*_.-r4CP*-r*' ivc --if> cv c_ a c* cn ^ ^ .

-i n ft r>

rt a. m - f f 1

1 -. 'i a - 1"

1 rt - J -*

- |V _ rt ,

C-m40>irtrv

v*. if m C P- *

4 c f <_. a

C rt 1 m o

rt n 3

^ -r^'

< -rt - rt o

i r- -- j*

_ :*-

or-

-4

t

rt N

rt f> 1-

J vj 11

? C IN

Vl. 1, rt

i^.rtrtU.1--

11. J I

1-1 4 I.vciNrtf

M4p*mp-_.-

._, 1. 0 n 3 ft ;p -1

r--.i ** P- rt ^-. .*

-I-

a. -. 4 a. 4 0

Jaava^J-*

- r-\t <_

* r r (\ * h 1-1

-. - - X m

<-"

TT -3

r 0 . a

Ub. U k.

1. ~ r #-ir&a.

d -. -rt rt 1 rt

c_ . tr.

J4 Z !. p (.

_* -rt

r

"*

r

II tj -rt "

C C.

X N n J 0 0 /l Pt- u. tfi if 4

'J

3. CN u - r* C O c

_3 c C <rt

*t

O rt ''1 N c O

U a. u u.

Pv_. "3.-2- Tig

Oil**

*P vCIf- C _T

cue*ftcN*.

CU vO v0

ir m m m

4 rt rt rt ft tr. m m mmm

vT CN O

C * rt

ojoim* m r-

m tr a m rt m m

rt rt CN -< m ^ rt art

9m <f> rt

fr.rt CNfcf

itot

cn mm cn

n * m rt in m rt i

lO^CO1

cr m e m

P- 4 rt IN

nFi

4 4

m m 4 rt C CN < rt m cn m IN IN 4 IN rt

rtrt^

t * *P

om ii -rt

mc c 4

a

m ir in rt

m o a* -rt

cr m a:

_n CNCN

1A DOt

rfOt*

j9 m * o

CO C C 4 a r- rt rt

ir. - ft inm i

!A iA 4 * o * 3 Nmatv

fr p in ccoff a.4<ro

is^rti <o-tr-r>c

: IN C rt rt 4 P-

a _

i WIN IN CT NT 4 IN

b (N '0 (N OrtCT*rt 1*1 dD"l B

Iftiflv mmiNm wi m * -

IN m IN 4 CN

JI1'1> 3 O 3

p(vCf* a*ff P

ft CO 'J* -

rtO" r p-

ff Flirt n

UP m IN rt

3 o y a^o^o r^vomot

ff (vp * cNirmsc- r-c 4

Ml <6 4 CN

*0 N J>1i O i<l CN * Ift iliftrnQV

p- m a- a in a: m ma.44 m o p> m

o* mew 4 r-ttom r- * m cn 4 p> m oart) -rt

ift o jo m Mrtjfl fipv>i.-_ rno oin

ccpia co m4CN 0- cc ac- - m m ir. cr

-a r- m r- <m o m rt in *. inva) a-l -rt

ft 1- r- o* -n o _> o o* ot o- 4 * 4co-44 tr. in rt m phwb a. m m rt

cr rt cn ir> t*\ fi ^ rt40a-i p- 4 rt rt

*a]-rtrt -4 .rt rt -rt -rt

^ IN

ONOOl fMP HP1 O- ft <4> 4 Dh# rt -N

in 4 o ir. ** uiw a^-irm

4 m cr rt rtm*^ rt cn rt rt r- - rt rt rt cn

-rtj 01

m -* o rftr-iN jo ot r* o ift # o rt -i -rt rt

<r m c r- < *ir a a - c rt 4 m * cv cn

rt or. a. iTvur-m c cn m - * cv m ** *-. rt - -

-* rrt) rN rt ;n CN

-rt -rt rt -rt

IN art m IN

.ftirtao- a n n -NO-ao> o -fi ^ -nj

Grt>4r>0> tfcr-a <r ** rv m in it cn *- .- .-

_ rv rt mm

COlTa-l f C P IN IT-P4a- m*C4IN

.NJ rt rt -rt rt

CN O f

4 C CN CN

m 4 -* cn

p* p- 4 m

m cr m m

v ^ tti e* cr *r *

vCrtmr- m itirt

4 m m - t o

4 m 4 art 41

4 4 fl CN in 4 O CN

ft vt 4

fC4rt 4 rt 4 CN

O IN

rt rti

m 71 r r-

P"rr *m

nn mi n-i _ j0-

' noctroi, c _ ii

U<H*rcfii u rt , .r- a- i- 1 o- 4 < ib vc o in cn

> ft .0 O ft .ft O ft -3

o* vr m 4 r- o u

,rt4 INU.C ClPINrt

f PJ"

IP Hi CN O a

* 1.1

u-I

f n * 5 ,.i i f d

"CfSo.rwO'n*. *r

;>..->,....li < Ifi O'ltOCw11

.1-

ut ...-

'i < P* O If tr- '

-4 * a a'W*

P <F Q iP || (>. _*

N

p

t- * rt rt

, r rtrtrrt.-_.:4>4rt.Ni - -

(ni <*. 4tn<<c*f)'*>rtCr mr

CNWC iTTCNI-fCNlfllN

_ rt rt rt rt-

-1 a -3

f O. CN f 1 c * 1 c

t n m 0 -n rt-vJj."rt)rt'Ji7v3rtO^Ni 00:* crto-N-J*

u. * c 0 m rt -rt rtho-

<o - h r- l'- 'f m ul cn1- f - a. 4 o- * i*.

vi > j ->. (-.-i,****

i-l 4 a- CN rt trU'

cn p* o f. c*"

f f f* c a +i firtCN*4CN

t -* rt M * -rt -rt - jm-* r-art j.m <"*-

- fN |N m c a **:

mm m rt

m o m

-J rt C

U a

-I rt rt 0 vl "3 Jt 4 rt

U 4 a. CN p C m N- 4 s

j a, 0t: rti m 4 vf >c < r* K-

jU -rt - : "N

c

-1 vlc- l- D

-^ rt r I r u 0 3 rt J> 3t- O C 4 <t l: m P* O a a

U ' '3

43 bl If) CN 4 C m vt>

rt -3 N - C rt J

tr, p-rtrtt-unr-<N

c f . rt

NNrtPvXOCrt-3

moV'Ca.vfart<r rt-

rt rt rt rt a.

j 0 .1 t A n 3 r- .71

r r --

x J t O

l*'tt w c -- p.o!rN

-.NO

rt r*

c ir a

O -rt .- rt

* O + .

O

<! C rt -- .

trt JO'rt-rt-rt t-rt F^ft

i^i;<u crvi. n - c r- O1

n r J b r*'-rt -rt c _> ni m

ttp. C4|njl ;aP-vr

\l\ * 1 -..> v , ^

- if r-

a U u _. <f rt vt C P ..

:pp "ft. v.

'J -rt

-.

0

c i"

-^

m

rt

a CN c

-t

4

-

Nl

r.

Q --

^^ ^

*-vl

C H if

n

P- 1*1 a. O

m

4

,N) IVJ

m

.NJ

a hIN V t tr vc J 7

rt -rt

.1. m -)

rtrt C4*rtC_p c fvccr-r"*f. cu (ncv

# n -r^ o n o -* j-i o a -*

o a f vc 4 cv .-. o vt u v. .

cc tr. r- c a. r ii1 * *o- fi it* v o f c rtrt.

w.

'

^ rt -rt O P-.PUO-rtOr* oONft

.rtuwC_.U.rt4

J.0H.INOf*il.O-r44

1 h o>irt p ft 1- 1*

X ** c -f

1 X J

u u.

- a c.

1 Jr

k a. c

u -J

i _t -i o*

-, CVJ1

*f r:

1 x OUb.

3 O rt x -n rt P ift 3

wic r**-JC*ctr

aaa-rt

1 rt vt vccmmxiNC

1-* .-j m

loovpCNO-nrj

/ o c m o o p*-

!_*-

c vf^ c w m 4 c-

1

1/. -0

x i;

rrtsn-rtrt.rt Fa-*

mo p. u-i*v. mrtO.

-rt -4

tt.*rp*a*-4 4if.__.

o r j

If. -C

.j J

c c

o r *p* o ** J!

OaO<(j-.Oft4U-t--.

ji a aj

r- VC HftVOftlflH

rt (t # Xr-

-3

u ob. u.

p- rt o

F-a* o c r- 1

v. __ c

af. .N

ftrtrt li*'*1l f 0

fl itiiit-kJ-Civi aiOl n*i- - .

> --< > j -^l

w-lv-lV- wi\fn*i d.

*i u m --< * 1 _>-

o: o

J % t

C. L.. rt > - -

<roio-~<mfim>3. im a

rt ^ m ^ jr> _.

.-.:.

-3 J in 3

r, 5.a18-

< m oi-n f #

r**!*** *rt cr cn

0>rt|Art * rt

oi oi m a m ft >* e

tn m a? m ^cmm

4CNCN

Ci rtOt o op rt m

c cn in cc m mot

vd o* vo o vo /. in <_

cn < cr ft m in cno*

.

* m*-

cn o 4 m o

oi it rt o r* p- p-

m^ m r i-cDic-*

tftp-ocsr* r* 4 m 4 4 m IN

rt <o *rt o ** -ft p*

fSP-P-CL04P"

in **- in mcr -tin

rt rt IN) rt * o m

P Flltl.FI

m in rt co

(^ rt Fl f- CNrtCN4

rtrtrtin in p o o*

m m -" rttcicm

mn t m4

^> m in m 4 cn

tmrvo m m f rt-< -> a-

i _ftp- p- p. -o o ."N;

fi art If VC

49 ft JO D ft ft o i> ft

t r *. vC m 4 c c b.

jC rt 4 CN W O m [N ?-

-4 .U-N -H

1/,j*

>- e

b. &H 1 1-3 -1 OA F rt rt

J C U CN ft C *. <_. 4 2

t m Ci 4 vC IC C

rt *"' i ."N

c-1

Ul)

p- nn F N 1- j a * rt 3f 4 vC tj m o

'Jif. O) 4 if C *- |f> v.

N rt c -n -3

a _

a ^ C3 a *l < 'N m ifi 3

- cc c P- t- x c 4V tT 4.

?. -1

rt V& *c c m m in c.4 m

rt

XH

c Li'

O II SF UNO <r* 0-

C f |v O p

- p> CO . - m 4 v_'

> i -ft 3 ft N

... m o- 4 m

p- m *- I

r* cc -> cv

C m rt m cn

r o c rt 3

a. it a a *-- -rt

-rt -rt vfM

.j rt a 4 o

at ^ rt -* - rt

r f , m in cn

</} n 3 rt

?-. C f O* m

Or- m rt

n

- 1 3 ft TJ.Q iri o P, U. f. C

a rt -j

vj*

F 4 rt

m --. a.

-<

4 ir. a

- -. ^ rt c

Hg J .

**pia -sc

p -o -f r

at m 4 ai ii

_i *

rt |N vG o. m t O' CN CN

CN _T a

r f -T* J X N

or. ri 4 in ft

_U r rt -rt . 3 -

rt m fi m rt > in a.

e m # J

r^ __ rt-

: -

FIGr5"

-3 %

oi <0 (Ni oi rtp-3 r ft 'ft n ft i. _O- r- ,-\j

p 4

it- m * > c o i* in c cn cn m cr * no ~ r-0*00

a- CNFl rt

-a a w - - a - a

-149 rt CN r- f rt rt m IN CN art * C C O IN O * If C 4 4 <C miN CN * O rt

lf| (N UB m MS 93 rt * o

0>c a cr r* ai iftcc

o> in* m rt * rtm <fi

'Oi-ft oi m en fltrt jih r p. ft ot

a C vC v* I*. 4 C O* O- CNFi CN

I -

art 4CNvrt rt C CN O

;* pi m jd 4 - 4 a 4 o <z> m 'r*n f j o o ftinrtrt4 a m rt cn <r m a p> mm cn c mpn

t * rt a o o

m m rt art rt o

mmm *ciO > o cnco

P-3-OM rtinOP rt rt ift 1ftr m m r* m 4 o - am m a

.r m m m ->*;** omm mm m rt ooo

ODfOIrt O.v0l*>m 4 - yii* rtm<r

<tiNu*

MOOO OtaPA Ort-iri

prt- cnj om o 4jir>r*fiF. icno n 4 cn 4

cnooo a-* m ov f> <r m m

*ooo cn cn m a> m 4

*tQO?V|J3 rt -rt vO O CN)rtvO 0> B ft ft Pi Ot O i)4 cr cn r* o cc- rt rt rt r- rt m <-* -* 4. c* 4 cd cn

O CN IN Om*rtc9 cn rP* *CN CN

ft n 0 firt0 rt rt

CN

a. cn m m i IN m rt i 4 4 rt ooo

* r* m-

cn rto

Acnooo in rt a. m

m 4 rt cn

CN O

p- a m a-*oc p-* p- 4

mOi-TiFl If. O" CO m 4 ft cn r**ocrm m4 cncn F PFIft

ft 4 ft 3* o

CN O O O

ff nf r- m m a, o.

m r j\ m r- ft ort ic rt m 4 cn f

C p* t it a -

- * .1- O P* r<4 T

or- - m o- - - .

'

iTCfP- COFir-

r ft o 0 c h j a uj o f ^ritcu -mcrb.C4r-i

-_j.a

(N rt rt 4 V. f, Ifl (N vC, rt (B P> <

N rt jv ^ ^

vt c m rt

ii. CN) CN 1

vC O -

- -j ,-. rr

tft C

M U. Q,

-t3trn*_aN."4*

."if.rT-ur*mi(i

; -}*la3: O.CGa-4 3 i m |T, HI J. .N rt : N rt .rt

P iN JO

tQ rt tf. l- |NP-CNrtb.CUmNrt

J CN j. - J -

J C O-

a. It -T Fl CN |N4P-P-Frt>tOlrtc-v-;CN|NO.

... N rt) am .-rt .N J-a-a

O L. u.

** m ivi

4 (U IN Ot

4 e o

ft -rt J O

a f 4 cn

mooo

r* o

4

m yi n *>rt_.Fi

m4_Ta- vtr-OP

cn ir vc i ir. rt 4. m- a n

r ni

O O Cj

rt T rt N -T _)

u. o fi m vc a. i

aU'frt-crc r.n _r

-k-4 .14

^Nrrt.jr7\ro^-5'Nr--3>.c'- ^ ) ji m-rt^s

Pc-iftfiNb.u.iNO'irtum r-t iTiLvVv.ft CCrO. f a. m m

r- - c c C C F> U P-aj r*

"3 N -rt

C-

N

ot C. IN m if rw

'IJ.IOti JO^ijOB.HC- O 4 O*

J*Omn.oOl4Mlvl C - if CV If .

7 ' '. i M i i uj i i ~

t->Cin4rnCrtCNm|NCl4rtCNIN FvCCNrt p-a-.OC OOt,' *-.<_<

m'Ilaau

- O J ft

) JP IN p* m

u;_#

J-*.* 1 4 a # c r-.rtvr-rt)-rtj^ift o

J o< rt vc vt. rt rt cn aw4ft + mmif. c

^^..O a

incirtCFiM\: u, i m cn rt a *-c c

P J) 4frt *

(1, va m Fl

C rt c

OC.-M.4 It <t 0 u.

C' 4 HI cn m c m CN F. Vt

jo *. rt J* r- ft

0'0>rrtiNi^tr*

,

0.rtF>rtCrt4,-*f' > a cn

V g- w. if.

o Jr K 6

JmJci-V" **'>Sm-rt f, oS o o - , ..

h . . .li~>'*o*a.mai-ciFvcc-a_i

n W

- e

r) ft r. fi o

IC 'i > I

rt n i-

a. **. in

OJ N -rt .

O Crt -rt O -rt ft rt'TV'T,

a. rt a i- e* *c a /

vi <\a p 4 ft m _,mv**"i-t'

^Jl ^ rt J v* .J 4

a j. rt .-vj -cp ft

a a

ft j r* 4 _JJl -H 1 3 nm oji jr*-Jfin

hi - 5 o 3 3 -;- a P 2 2i-C iT o S uu. u. - "- u

JB O -rt 4 ^o.N-rtrt._;-*ftn

P> O1 O14 (rirtu.tr> m m -

rt O CP-r^CNrtCO'vC

* -* . rt rt -rt

c 10 (A

X Cm C C

\ O N rt ft trtf^.-rt*|rtO T 3

m m o- 4 q. a ip <vm fi w r

/_- i -_!

U'- rt c aCrtCN4n.<-p'p*

_> -rt ii U 4 rt -rt ^JJ #

rt 1.3 </.

X J J

> a a a

o m sr-4*

_: r3 i i o n

r m o o> 4 t- :icrp-p-FWE<r<ro

a.rtiNotirtrtmvr Mrta c

'-=\%

o 0 0 o >

0am rN '

pFtFI W V -w aj -a

1 .

-

,* ^ m rtrt rto

2rt2__? -SIT!-.

*-**< oooa O4mot ao.no* merer cn m c* a~ in m r> - cr m

4 it -.m

4 CN 4 f>

a

4 m in a-< o no mooo

nfior* m in m

rtCN rt

o m 9 pv> ft 4

m 4 m ce 4 c

cr 4 ^ o *-

.(AN-rt*** 3 vOOl JtCB

Ic^ri cv o-* omr-cN

, , # #

t m ai rt o

p* ov m o *rf a 9Nh s iiftoin r o m ^ ,n<m

NO Ft tCCN^TCN .ft rt 4 m CNrt rt CN C 4 ir. m

a

vrttnerf) 9nFi ^ m m c in* rt rtrtoo mooo p- ^ Ft CNO CN CN artrt rt

rtlfl* it P rNP*-

'r,** *cn nt mr^cN

, i-

i Fitrtommmrt oooo.

5-T _?_!._? ->OvlF| ftrt4-a Ol ft -31 O OvO O Pa * * Ptoomo nco cc m cn m rt 4 m m m at rt M m 4

a a aCNCNcrm w cn rt m rt in **rt m mooo l w*~4 rt ma

ri 4> f* ft ft (HA O A

,*p *IP^rtlO CN 4 Ol CN

,

,ftin<rtiNTrt OOO CN O O O

*J ca c* 9 ommoc ** *it - pi o. r* p in o <o * ot m mrt o> > r- c m m p * cn 4 m o 4 rt rt r* cn <c 4"* a a a a a a

CN rt O, m Mrtf rt OCN rt CN IN rt rtrtOG CNOCO r rt vc 4 '

liM* o m -rt) o -rti^o-rt

10 (TvO P C P- rt 04SDCN

IM fir* f* Irt 4 CN

<V X ft * -4 rt ft JO (V| Pi 1ft ft *o P P- it O"

o ce oc> cn r-4 cn * p> m 1- c p- o> m * -.

cr f> p- m rt O CNOCO <_. a-icv cv a if cn rt

3 J * rt39i 3rt0.it

ciro ****** crmr-a-

infrt CM 4 4 IN

rrirtOJ in

cr ce 1- c

p*o miv-Jifrt r it 3 o oto*ti <n f v ov m ot-A o p-oit-o m -v p* 3* o> o -0 <o m -a ** - ft o it _ m -rt

m m a-v cn a-> i< rt m 4 Oi m m vc o> m o. 4 p- ps aFoi rt m m 4 rt m a. c r- ir. r- a pcvmir a cn a cd cn vc

cno o o vc r *r *r ooi 4m ->.p- m cn * rt oco.m i-fcn m cn m o> a mno c

*N- rt -) -rt

p m o o * 3 - o*-O o ft '3> m ft r* 'ft /to ji 3 -* p< t/t m rvi ft ^ r 3 .- -* rN u o -n. r* 9 >o -* .< *n o o jni.ni

mp*rta-i c- ** cn crmcccr 4 c cn m m - cn r mai cn m 4 4 in m vc p- ^r ^ ^ tft -*j m e - oo-f. m c c

cnooo o 4 p- 4 o * m m 4 rt rt _. 4. tfift rt rt o o 000 cr -> rt m m a cn m 4 m 4 o. ic a. m ^*

o*rtoo. -rt4*o-ft-rfv(#-o ft r* r- -rt o .u 9 it n* p* 01 o ft o - _> 1.^011 o ni - 01 m 4 ft - . --* op-

- a a a

ftcco iremp*tr m m ir f. rt ... r- a m f. m c ** c o c *- c 4 rao.F. m44r c-ofc a-c

r4 *+ -rt -rt rrt

!* F tvi it ft rtI kit _tit4-l Artrt O-ODM T* O*

'M * r* 9 3 -rt O -N * ftP-~*."N O -ft O 3 3 fl ft O 3 art

nt cs in in m vc m c o- ** m o o - vt ciyCircN >_ c 4 if. - m oo>4 r- r *_? cmcir r-aap. 4

4oo m c * r* inxwj-pi mo-iNrt mo-aim 4cco.-

c c Fvrt4 p:-xp*m p*mvrcv r-aoc o.c

*l j. 0 d N rt O 3C ; ^ r f. o. u-*

.

' * -t 1 <

^ (S w U. 0) Fl CN rt

s*mcc<ir it w

' 0 -* X P. rti ,n 3r*a.vt icic

1 v

1:s

< *rt * nj 0

* . .3.***

f ac m m

p- c c

Jl -rt 3 O Fi A **l -* rt 3 F - a ft m ft J li ft O O .-N A N O ft 0 ft i> m ft 40 -rtrtOJlftOftN-rtJr-.OOJ- rt j

ar~ <r O- m 4 - CP p. r- b- 4 rt a- ir _ r u vc rt f 4 4 C p- cr a- 4 Ft P- vcr ocvC4u:miP*-*:pcccN P- P

m 0 0 cnvj. c>ic<->>-.mcirtcooirtrtU.iNmmFi mooo m c 0 r ii 1 ifi p a CNmi-ip-K>mcNb.rtO-|NC Ft

-rt U.N -V. J J rt *. * j

Ifj rt C a. *n J (_ c.

l - x - _u >rt) K -,

__> u. &. j b. a

3 ft rt -* 1 f rt N 3S <F -N * ft 0 o* 3 a ^1 -n p* r -# r # r H ft 0 /t 3 r~ 0 -rt 0 i^moft". 00 -ft 9 m ji

m * iT' f> U.CFtmcaoo r-0. yicmcNiftC C m rt o- 4 t m CN CN 4 U. rt "* CNCNaO 0-*ea i.*-4<tit.CFVM.i4'4 Vt rt

a g< p a> < r"mmpH4iNrtrtf--rtmcN vC O C m Q. CN 0> If) 1 O-vCF. Oi^OCCNrtrP-P-rtC 5 0

X *+ Z- * Hl -rt) J -rt rt) -U N) S N rt -rt -J

c Ukv1 (_ U -u

J 'S - - 0 3

C' *' -3 Cr-

-3 C

-rt 3 rt ft B l> m N -rt.v*33ft*r- ^ - F . Jl -rt rt N O rt) NJ P i*> 0 j rt rt 3rt3mjft N-i;rt rt-*D F O

(- olcnc^u;o-c rCNt_-*rtiftii''f 0- a. CN #rt c- 4 CN |N r-iC'I U vtO.1-U7vaCFIvCFa.rtt_.4P' 4 Fl

U"C*4 aiftCMt m cn in m rt rt ._: x m m in 0 - C. _. X art 4 UJ vC vt/F>c.r'if-^rt 4-Crt-jifiircNr If O

_J M * rt -rt ~3 N X N j :n -c -

t z 0 l>

-5 ^J

C< Oji -n 3 cnj n rt -ni --t -^ -rt -* ft 4 -rt 7* -rt 0 _. rt-Jtftm onj _> o-3-vi t-M.n ni-n i-N _f|

rt m < m cc a>m^> m a a. 4 CN 4 c m 0 piHif. PML4m(rhr"W7 r^v rt u;

t,FI4CNvC 4 m m a cr 41 cn -P- C f rt - P- P* CJ rt m r-mc^-Fi4Ocr-mc nmn h Nl O

-t m rt si m -N .- .-N rt

1- 3

x c *. ->

4 b) Cb.' -

ft ft -rt *

pota*

j-r* r .-vi -rt 0 r vN ft O -Or ft J*'N C U r f 1* J r* O "1 C 0 rt .-rt -rt -rt T r

cr Ft ui <r fi m rt vt UD CT rt If, UL W O* O" ? cr- r- cn m a. ui 4 * a> cm * u.oirt*v 4 a 0- 0 cff-a if) ? n cr

- r- in rt* c c-r 4 4 &mm4rti_.i

r-mrt- P> CN C C 4 a. m c rt 1 C C Ft 4 C

: ^"io il P o p- -tv

3 r*0 m m

c- n it

ao-

m

o-< -n -* o

o> cn a fi

u: rt rt

-r 3 .n*

m 0 4

r- cn m

oH 3 O

a. m o

c.)-_m ^

C |j. j <. 1

u. 1

J * 4 .- JO

iftt Tirt

J .

. 4 cr o* c 1

","* ft. m oj

6^nfin:

o in *

: in o cr '

t

: rt in o 1

?

l/J-*

-I

ex c

r -vlr* 3

f h a.o>p r

j J> t *- -* m i

.x x * -

*t =.

X 3Ufa-

*& tf m tn 1

J-. IN - rt m I

x -r

x P* O "N P-

1 m * p^ ft

SF IT CN C>

.3

-

3

a.cc<moo-*rtir*i

.-vi'rt sop-^ 71 P" M N

:mp*o<-.rta ,x o f r-. 10

.pv0-lrtOrtr-rtrtF>>C|f .;

cm i. rt

O *

5-3. 3\W

C

Jt 0r-f*

t -

c o

-3 ft >ft |rt-

o. f m rt

4 IN |N CN 4-TFtCV, 4 CN |N CN SftC C rtrt

ft O ft ^ *4 O -ft O

rtvc-o.it. <tcn p-

m o o <N]

-f>rt4 mmmm

Cftoo-it itotoo mm

at a. * p- mcNvcm nm

oooo oooo i- r-

TfirtJl

m cn m cn icmftcN i in m m m ft rt c

to m ft -

& vr- rt

o o c-rt m rt cn 4 m m m c oo o

co ojoo o 4 CN 4 CN 0VO4IN F, 4 CN rt O wrt O < , rt rt -.

m m a.

0*00 O O rt o c o

*p

rt r*

rm Fl IN

-

INC C O c m m art a-ir-4CN p i/> ci rt m m rt mrtoo >oo

co 4 m 1/tlM ^O O CD*1rt 1^ ^ ft -ft" p* ot m o rt m aft w c o r- cc

a a a a

r--a rt a-l CN 4 IC Ct 444m PIAftO i rt O rt o c o

CO IN CO p- 4 m rt

-. m

4 CN

OftftO O-tft ft rt 0>aDfNFtC 4Cil.CN O- vC CN

f. m 4 art m , CN c o c rt rt rt rt o

oc o c e c i-if!ciN vcamo.i cr o> i- m 4 i r- m m o rt rt C rt rt o o

OO

3-0 Ot

iO"" IN rt

wo o o 4 CN art t m 4 o cn

mot-rtm fti/to-*

4*-rt r- rt vc

cr r> m in m p* 4 m o o o o

r rt -rt r*.

4 F 4 CC

f a *o f . Ct F rt r- f> cn > 1 rt II C IT - C C cNrtce cc

O in 4 f F- CN rt 4 c m c c rt

* o dOiv<rI'H (N Ifi 4 rt

i ft jv o r ft i> * n

ooo

rtrt 3Nmjr*rtn o

ii. c o a. in a cn c rt r:

. w, a-im * rvt rt

-

fwee j* ui r- p- cn m c

c o

p. r.

# u ;nj n rt r

vT Ul P* P- 4

*j cn h a. m cn <r

iNmocrm rtirt

rt.--n .Nflf c n v P 1- r-

i rrtrt4 C - ifi if C- tr, v,.U'

4

J rt***- C

'

af.< J- CN If. % rt X *_ _ Cl

- tj -*

U L-i

o 3

Cj rt-4 ft 4* t3rvftin 0 .-*

.Nrt)

u rrtfN-4a-CN<rarcN vt f-r*

_*.

)_,

irn.i4:4rt!r'p-Ftrt vt CO

.3 rt < ->

rt s.

; 3

C; 15'_> o'rtji e vco hh 3

dO if. 4iNirtOvo:opm O' cn tc m

-4 C.

o r* <t> 4 rtiriNrt >*t o o

m vo <n m u c

Ik IN O. rt rt

* CN O. Crt

i m cn t- C 4 IN rt

r- o m cn t-

m c * u.

MH OT

n - # it0 "ic. o

I # B -N N -

n " j. r -j ft .ft n ft "3 -rt rt) r 3 c 0 "rt -rt N|

tT a rt - u.p- OV 4 - a cr fi m vt a. ot. ao

v.r- cn cn rt c m m o r- in 4 m in j m m cn in

-rt c .N - rt -i -c -rt

a % IS_j C_ 15

3 * o < n * -rt rt rt rt 4 p* t

r if in cn t- 4 IC IT t f m 4 * m vc cn -*

rt Oj O. rti o n m ir vr c in 4 in tr vc m in

>* i.J N -rt rvi -rt

X

X _ J

c tvJ F-

3 3 P- N J 3 <-* o -n e

a r* c 4P-

O O" vt IN 4 <x m in cn m

m 4 m m - p- c* ft b. c

a OFi rt O O 4 O O. 4 J

cO -vj rrt O O p* (* rt) O

'

p- m rt cv vc c m rt t- 4 C

Om rt o m rt o 1-

__! O -

"3 iN -

C

X

4 -*> r > o n h -o ->t -r

q> rt rt rt in vr c J 4 rv

*C

m rt o p* U*r.

o c P CN CNCCNplOC XUNVmClPO-rti cr rtiNrtC "fp - c 4 vr 4 cn ; Ft r* Fl CN

P*-rt rt rt| _

* ft0

.*

IN vl'

^-c c

rt

a r o CN

II .j*"

v.

**

a

vi

'"-*rt * O IV)

lp- n,

3 '3

tu. '--cc.)

__,

3. t <

31

trt) m "3 -rt *

ort O P^ '

r. cv i- i

o-rt :< rt O ;

(i. 4 O vt) <

INO.rta-FCN3_C f-40.1 .tP-C. -OltOCC

O r* Jt rt

I Fl it 4 |ft I

cI . o rt

a. m o* m

CN4Flrtrtit.imlN*4<rO>rt

Vt. rt

-

3P-

>*.no o

m cm in i_. cra .3

rt rt e a *

Jt N n j

tm< hit

a a U J

c c< * -

X c

-4. iJ

X JOb-

ft ^ O .1 -rt C c n

o a vt a

-rt

4 m p-

vt c ap- vt'

.>i

w

c

N

f * Oi rt

4p- a P a

jr* rt rt rt Prt r* O m

Ni ft vt SO

crt -rt rt

4 a Q I

*rt

m in a -

j

w_

p" rt -3 + NJ -o .-N! rt _, ^,

vt 4 F. cr r m 4 a. F p

1 -*

c Ui

.3

a c O

ii

a

cr

o

to

J

c

4

rt -

o -rt -3p 41 3 4"

J ft

4 ft 4i"

rt P - m *- a. m c

.3 %j jO O Q. o> r rt m <

X C *rt

.< c _.)

o c 3oe ou.

^^ .5

-o fl-rtit^ moooi

"a a *

,^aj?WtlftFtFl lOFIrtO

ObOrtiA ift Ol 03

aDrtTr> FiMClrt

OOOO oooo

tm m o ft p cn)mnmr ioo

a

r* * cn 4 m cn

c- o-

ma a a

CN Ft

<0 4 (ft 0D P PC ft AlAOmo. vc cvm 4 cn c cn4 m

a a

<Crt(Nrt rtrtQO rt0

O <D m vN KlrtiAO > CN) Pi it F 4

m io m cn r* cnp- o 4 m a- t*

a a a

m.44rt a*ne ccnoo wc

a

ft m oft it io o co m ftrt cn ov m rtift t> m m r^ <c

oooo oooo * r- m

^ o m jt o* o* mmp-ir 4 oi m 4m

rt rt O rtOOO cc m m in ) rt4 rt cr cn O CN if rt

n -n * *r -apcr**n

OA CD art lOUIFICB

<MfrtO rt ft "N .4

pc r* ift in

0*aOO art O O C r CN CN

o m m -rt * <4<o .

o. r* *. cn mircri

m rt m a-. ft FtO rt CN rt *^ P- m CN rt 4 CN 0 CN

rt; rt O O rtCCO CN f* CN Ft P- CN IN4 0 IN IN P- CN 4 F m rt rt p-

4 o

rt rt O rt rt mr* rt rtm n in 4 4 4

Vt F 4 4

F VC CN CN

It .> p m 4 m p* ir m o

p* cnj o. m chomm

BIOCP 4 C CN

rt art ._, rt

** P 4> *N ft ft ft ft 4^t<N"N *rtmrta) fN) 3 ft P CN OI <ft <rt

4 P- m C ftmC4 4 C 0 vC <4 04vCCN 4 0v 4 m

in 4 rtm r- * . crmia-rt r- vc - m -* o o m m c cn vc 4 4 4

*! ft mo -4 rt * Ot Ot

>* P" O** vC vo 0 *-i eo

BO p in ift m p- * m o

IH

'P OO It rt) ft * o o1 * 'if ar P" cn 0 0

;c 4 m *< in c

ft as * 4

C rtFt rt

CN rt O rt O > ifi ic rt r* vc rt o

* f* oi ft m ft p * ft4D<*oi

~i>r*-4 m m p cn o* cr 4 Ft

rt in o o o m 4 m vcomot - 4 1

c c

*om--i *. 01 r

o> tc m m cn I 4 4

IN rt >*

vt 4 tr c If 1 O CNOO

*N 4 CN - rtif* m f O*

.*/) n- jt ft

ff. u%

c vo m

j ft 3

_; m o

m o

,3^-p-3s Q (r f, m vt

1 Hi*i>i *>

'"i^mct

c

c o rta. vt <

a

uX 3 0 "

-t IP vC C

*Co^M 0 w'mr

^ J 3 m a n* ^ * a in

"

C * * in jN

; o - -1 ^

J * :

ZT* -% & r*%

(T Ft vC 4

CN

m m o *t

rt I- iC Ft b-

vC C iT- rt

m rt r b.

I- rt

.i m o n

O 0 IN IN)

4 C O O

IN CN m IN

rt , a- O

> 3 O O O 3 flfllflrtl" ft v] 3 p- i m -X N N ft rt >J>-rt "SI Frt-rtr ftrt-D.r-QNftP-aU'rt.rt.3

- P- 4 C * O-mmbm4lTiCrt0FV|frt o- rt,a- *-. IN P tr, %T p-CNrf"'*". rtCC4CNbOU>OvCa_;rtrvCCN

a._Ja.a *-*

lAINrtlA.r-vaVINrti-taCNPvrtb m m -* c , rt a^ C m 0 0 rtCNif. rt|fjJb4vC4rtrtO vCrtUmo-m

4. -N -v. 3 -rt -rt J .NJ ta J.r C ar .-_;_;__.

-i X -b

fa A. -: b. Q

r r- j. n o j. on t -rt-rtrj-virnioo ft N "F O P- 'NJ ># -ioonrti3no-*r rtT3<is"NP-rt*

C C. t U f 4 m r* >- r- 0 t C00*cn v(. C O. m 4 r- fi m u.>cm>a.macN*cc4Z writrFVrtCmfiiN

3 *-;*.---_ a -. a

1 vf .NC O *r- Firtiimo.rtO.-HC'Of-rtC vC -n 0 co <; vc.-n 4 4r-F To<-|-rviPr

rt T 'N rrt rt J ..-art.__.-% O-^

c O b. 0 0 u:

rtj .1 3 1 -rt J 3r- r- c c i- -. c-

T it 3-3.noi_C-3T P > 3 "F 0 0 n n rt-rtN*-' JOOr-*jrtlrtTft P-rt

f rt If b *C u - a cP^wma.irmoa' C f. F)a- p m m fc-c -rmf'j.r*iNiNp--r -rta-c r a o-

if vf

in 4 P. cmoimjm4a-G 0 C c m c c it U. O a. r* F. C 0 m P-vN --*rtg-C_JCNvi.CN

J N rt "3 N trt J N| < -vj

a : 2

.. 3

c a * v

t _ 3C' c a. *_i c 0

Irt-IO'fflPNrtJ "rt/t-NI^'VI'VlftO rt "ft N p* r^ 0* "O n-rooxro-ft-rt 5!,rtT3-t:'*,'*Nrt a if. * C PINaiMiftPlftf 4 P> J 4 rt a) 4 m JO-O 4artf-*mm*-ICN_l C-HUXf-ifiW.-!.'

J

m4o,o,rt*rFiv mvooirt O C r- c UP--np***c.rt"ftrp-. inof>o omc

n ;-* -n rt. m 4 T rt -rt rt.

t- 3 *- ^ c- X

v '3 t c u a

r__ -i t _. r w

-rt a "4 Ti - * rt .-N ft -0 N*-,:_rtjpi--'4 j t rt p r- . -vi iv

n m m i/* 4 FlO*r CP-vtC CPI^O 4-vf

P- 4 4 in 0* m m onco Ortinm co-om wcn^

. CN 4 Ft C 4 4iniNC-t/'mF)f_or*o-p.' c rt a- C4mr-4CP,*',ccN_ 01 C C P. 4 O*

C* f 3

a o *r

* N O 'N rtl^rt-rt-N-Nft-

r m 0 c f.

< or 01 r- p>

4 c 1

WO rt.3

P-vC I

1 m " t * m vc f. a -tm

t m 'J "

FvmaiNacN'V'ip r - *

^m;i*ttir*p"iflh

3 j

a: cr e -

(X Ft 0> p-vi. U.

-JJ

Irt m cr *; 4 l-

m c

.3

T. a

fa. fa.

ji O 3 3f-

F -4. 4. rt 0. F| m I

a ii *

ii m vc m - 1/. if1 1

m t; *>

- rt j "3 rt 4 (i 0 1 TV -_, r r- N

0* a If) If 0 P ;) ft. c vO vt

rt

u a O ir (T 4 vf ? rt 3 0- If!

3 11 . * /1

-J A fj

X * J j. a C c_

P* r rt P- 1) j -F 3 r P 3I-

va CO 4 -r *- n ri >i r il cr. CLP" m

_ _> j a

a cr O 0

c

c

n

ir 4 ^

ft rt

4

r * rt _a -rt n n

if. IN a T rrU'

. rt ft

J

i/;

t

vt>

f.

rt 3 m n rt p-

fM cr. "t r-

u

-

in e rt

c=>

3

4

CN vp

c e e %

Ku tr- % 5|2

c

rt m it O O 4 -ft

t ft o o> a r- c

*0 op* 3f* 4 CN CN

,. P.OOO rt O

a* a a a a a a a

) CN rt VC rt If CN O^ m 4 art CNmiNrt artCOC

m to -n 4)

m q mm

a

rt c c o

m o ^ or

CfCN Of.

a

*. ct m ** a- m m o m OOOO rt <

m ft o* ao p ^ r* o* jt r* m m r*.*n n i >o m m 4 ^ot*v_oi

a m rtm ina>< a rt 4 m p* m rt -tn oo a> in m

IN oo O CN o c o <ft m CN rt IC rt If. CN r rt y 4 t ai o cncco irt cn m m4

c m 4

<rt c OOOO CN OO C

rtO O o moo tC IN*

-)-*r- cn w* tci t * * oooo

rtoco moo O4 4 * m 4 m ~**r f* o m rt oooo

IN iC rt

m< m .ft rt cnr- m m ^ * a- - irt OOO mOOC

c e mc c o IV iC- 4 4 cncn m rt CN IT f rt a o moot

: f 4 4 CN 4 CN art fj 4 P* -. OOO m00 mCNCNrt m F* 4 rt 4 4 rt CN l IN CNOOO Ft G <

rt O m m cn in

Ol ft ft O ift it 3 m rt iM

** pi * ^44Fl Ov Fl CN P

rt rt CN CNOCO CNCOC

a a 4 4 i vc m cn rt in rt in

0 c it r-

C F. CN C r> vr 4 c tn ir. cn cv rt

c cc m i/. 4 m tr c in *

CNININ F. C CNOCC

9 o rt -o

4 V. F CN

I O O O CN c c o m vo 4 - 40<o> 4000 CNOCC

titrtoc mcc

/IP* * p. rt o n irt) m 3 n -4 nP* IftCN vt or if. c u. cr m vl

F> n vc vt

3 t,v c a rt

Oi -rt

O

4 rt CN *C_"

r

N m rt frt o

CV !_ vi O 1- 0. 4 IT- tr _i. C G P"

c c e 0 1*: r

"3 -4

CX

4 r* CN C V 1/1 If)

a

0 P-O rt

"1 OF, rt- m n n

a.

it -rt rr* <-

m 4 -J CN * . 0. k- - m m

WCNC C r* c o*

Om

rt CN)

HJC

4 f. CN_:

p-ir m

rt N

'X

*O 31 n n * D

Cf cr ir vi

0 b.

M"t

OJ ft ft

4 m rt

r

4 m uu

rt -F U ft J* N 4

rt CN fcj IN4" (N VC

MiXftFimfr--*'<rcm 4o

4lfir..rt^Jrtrtf mooo

cn o o m cn I'CMf

rao-cc-ntc co

>* J O 3 >* C

m co *- r- at i

.*. .**'-" "'

C I* m 4 m c tf. 4m&*v0H)CTm-art*.- !

i**o.

C co trt m -rt) * r _t

_-

T *- T* frt"

".v( o- 0 4 m o P- r. tr .i. O m a> o* art

rt a

3

* c c; Ct c IN If' m a c ii. irt 4 OC !" a.

ll Om J i".

to if If.

>* _J

0- aN U 3 4 art *

st -U t- a. 4 vc

D. 0XJ -vl s -U

"* f 4 Irt rt o. m F -, CN) m - 04 01

3-

Ooc

x xm t. m X. o.t .3

* 3Oh.

>3

b.

3fa.

v. ;vj rt -ft J -ft v_ ft TV j-

,r c ' r pFiniN

fv- a. cn CN a <r vi

f irt a

rt rt m (L

M ..J --"-'

i-.0i0imm-.C'mmrta.i-.--ooao

> o ,n a. r- ft jf fti,,*o^-'irt.-rtr-uo

Fl rt fN rttOCrt CVIfFI*b*TOP*a-^^tJ

COO 4v4.-rtFvVj0m4CNKvrtmrtat-.4

J -rt .v. 3

K -J C O.

J -* X ^

J fa. Q.

* rt) a ^1 Jv ft 3 "3 "JrtJrtm -It r- Jt 3 7.**

a-< rt iLtftplrtlNUvrtfilCFIc >t PI 'w 4) U Hi

-j. a. -3 -

c aOP'rtrtT 0-^FlFt^rtr*0**rtfr-*C

a. rt - n i-. *_*

c? o b:

_=- .1 3

C *"J C-

r o P rtP> U-OJV J v| | -M-OJ 0*3 rtP* Kft

HOCvT-t-bOvivCOC -1 lfl*a-

jtTa.

(jV Vt CL CN VT w vf NlNFt 4CNa-0_-;FI

-Jrt X-N J-rt <^

c a -r si

_:_,<??

TJ'3lft-V-t',-a,^rtl J/t N-rtf*^-*

j^if. oi4t-if tn4ammvC4rtm

^j . aal a a j) a

oL,0P-4mC,P'^*rt4O4Part 0

- NJ -J ^4 N* rt ^t

H rt H X

1. t IV. J

u r? Uf ^

> O It N N CP^-rt'rt-NJO rt O -rt CO

(TV4P-4 CN<t'4 0. 4mm 06

U.aiaU- --__ !_

c f-U/vC_PU.C0*4O4C if, * irt ft C

.<N "- -N

tf, U. IT

K X OB

o c o oMmOft'/-*rt-rtrtrt)ft0O-*rrt)T

a.r*CNv/ U C-.F- ltlfiO.Q.4Ca-.art(_.|N

-a a-rt -rt--

o,O' li < * >i c.4 art m a. 0 cr 4 ncc <-

li J.vj _Jrt sJ .N O NJ

ia u. v. m

a) u J - J

c o o o

o r> -i p>.-rtmo- 3-ftXts> ^

o.p-rtirrtrp'omorp-p-'Oam

O ~-l * J J ' *!

HlCj|j/vr'4rtC-*iNmi-ia4mirtir

,t n i_ f* j6 "t x ot

_. => 3 =

xc* a a ^

Uh. u U. fa.

4CCU INC-c

vf m rt mooo

vA rt O O <f C Oi

rt r^ o -rt *N o 4 o .rt Nl

"" ft vt 4-

->

m *" Fl lf> p-vC

4 CN C rt vC rt o o *- r- c c

j

m -* a vroooa a> o

>

Rt5^ &!&

J0I-4P-

Cfflnf-

**> * *o

CtflO P* it O 4* Ol O

o m cn p* m ot in m m rtm

_ _

rt o oooo c > 4 m rt m rt < O C O art , f r* 4 *rt

it ot s r -rt CN) 91 O ft CO -0 < ft -rt

CCin iCmCNrt rt V43 C CN

. a e

|>rtO OOO CNO i m4 in a> vt OOOO IN m CN CN

. ter^nap) ^op3 cnj

*! mFioi o m vo cr op

, a 4

M * * in Crt NT rt IN 0. trt OOOO CNOCO OOO Ft O m cn m

'art rt OOO mooo oooo

CN m ?* rt i*a -! moo I 4 rt o< w. m rt rt ooco

CN O r- m -

o>po m in cn cn CN 4 m m cn pfliirtiv m CD CN vf 4 rt rt CNC-OO CNCOO c r* cn m 4 if CN 4 ai 4 iC. .

m m cn cv rt CN CN CN f. C 4 4 CN v IN IT-V4. IN . CN C C C F * IN F. F. P- v| O.

"IO 3 ft JN o o

'Oooo. m 0 in tf> r

f a

1 rt

If.

1 jio * e j_ p p- -. p- a o

rvirmicu it pf - .

1 t --_) . -

J rt j^ 3

.J c a.[* X- Q.

|. . rt jt rt -n -ft p> Jrt

'* d_ vt m o tu u a-

-rt

iNvmrni-^iNrt 1- vt

J

4 C

IN If, t*. ,N

IN c c c 0 vr: in m ir cn ifj

i n o n jo o ft j j\ r

a_r>rtrtmv>rt(z.it.o

O4" r- c* *

- __ aUNFiu. np-(s->

a.

_; _k

Loo*ilrt--;ftrt3r-

_.r-rttf.u-v-.---ww.'4

CNic*iN|rtinc*rN

a in ftp) ul _. Jim u rtu.rtrtFi

'J'J ^ 0 3 oo rt

I\ I 4 13. . ,-

i\m 4 CN rt

C<-" s ft

vo a

ft ft O '- rt rt ft

it a U F" cv o. a

3f- OM -ft CONF

& cr. .-. j. c a. c a-c

_. a

CvmoiNrtiT I P'

rt m o o

rt >t o

4 C C C

rt N J rt 3 3 rt 0 .7 n f Jl J rt

_ St -_, W -. b_ N w_1-

1 _ V^ (V O. I-

"J _>_, J

ir 4 ft HI - 4 CN rt IN

V. -rt i N

c

rt rt rt

O-4

*-

u

-rt O O n r j n - - J N P--F

1- P- Fl vt O- b if r* r-v-. U 4 CN -

-ft

Hvnno a mmvc 4 ?

P -

* " "^ " "

r^flrt4C4INrt

%4A>-CON-)U

FcrarH-rttftp-0

a a -.

; * O m iT iT. ("

o* -rt

ft*.

In*. .

i 11-

'it*U

n J o -a 0o a< 4 m

la . a

* w in vd m

rvi

v.

X

on -ai o ft oa. a. 4 m

*t ct 0 ac 4 1

vj .v|

w

4O

3 v/t vO

"> Of I- O

n e of a>

_

pc c *

Crt irt ft

- a. IN

-rt l*> O -rt'

rt vt> It- ? *

'J * P* Ol 3 -C "rt '*# '

flvCvOmrtP-CO-rt'

rt a "3

GvCOrtrt CN 4 Ot

.'Ntr* . 4

j n 0 0 ft r. n -- .3

F. #. U. -t C. I

' O a- P- C 1 *F4 3. ** C if if* i C rvmoi""C 1 mm--..

r- o

Vt' - er

-* r- c

T v.j m L_ vO CD 4

Z 4 vC tT 4 I

C_-(Mmi

cam

Jm w 01

I J-rtjiN-rt-rt ift 'A J "*

.*o-c4rtcmcN

.4_/_

4 wDrtOLJCC

6 o O o -rtr-p-r7*.

r*0 m *- n f o* m 1

-a*a.3aaa*

. 14? C O tt CD '

I a r* - r "rt f rt

. Fl rt b. HI vl *-- !..

rtTIN'Oft 'I h

u-c-p-aaTrf^.

. ^ - j

3.H ttt f. |v r* O^n

U "rt > **

a o -p m -*

_ c rr, ^ m

. j a

_ II rt <T O .

F O> -NI O"

- r- if. aaja*<*

^0m^-r- inr-ir-

X rt

O rt ft O 11 1

it m ift in n r u t-

j_rtp"r-4m--*o-rfi

c rt 1.**

c _,cr rt c

,__ t - vu (V

J a . a .

rt a--<-<

o

/. tt.

- -) 1

J L _.

.1r o>0'0

l- I)l_'

CN rt III IL .;

J J a C

. 1 rt C> CN P- 4 - r

c C # v, r

'i 1_ :.

X 3 "J

0 CL I" art (ft C

C - VT IN rt tri

C 'rt

r.C 5-^ "/I 2

o

c p **

ftOCOft ift I* ft o

C f. cn m rv p ftFi

a

in o rt-cee

Pi ft * rt '^/na-urt >r-tirt *r ft ft r>

m o m m a r m ac p o air^ fi m *

a- a a a a a a a a a

OV <C rt CN 4 CN <C CN FIFiftC 4 <C rt O

) o 4 o 3* mm oIP-C 0 4 CN CN"

C - o O C

r- a. ti3 ma

mom m cn rrt 4 rt rt rt fv o c

HOP Cft 3 rt P- CN tN) ft CM

O rt ct vC 0 * cn m 0 r> rt m

(*ir-4rt art o oo m in -* wtifiFi

iNi-No-in ,-n m -* m

c a 4 m a 4 inm

IN IN m rt m 143 CN CNOC CNOOO firt IN m rt CM CNOCO

no * -* ift 4 ftcirap* r in * m

O C CN O O C m r- * 4 CN art m IN

3 > it 3B -rt 4 ^ ft P Oit rtCC art O CO QOrtrt O- CNCN CN

m mm o in mooo

"033ft rtmrtts o m

wrfto 4vt>o- mo

cd cn m rt - cn cn - mm-* -rt

CN rt m CN CN CN <

m o ft cn

ifi -N| -JO -rt ^m >ft rt rtP-

CN V434CNIN

artOOO art O o cn cn rt rt m i m f fv rt o o o

O fl rt CN rt rt tfl tt CN in m O Ft rt rt r- m tr.

o 4>

w c

O 4 at a c m m m

p-vcivm ircNom o- m cn rt < C C rt oo c 4 cc 4 !>. vc m cn m rt cn ' CN a-. o o

F, o o o IT. P* arta*.

rCNCN4 If rt IN CNOOO rtOOO cn 4 cn o o

r n r n

f. m p- m

CN C C a vt o- - 4 o in u; CN c c

vvlP 4

+ *rvirt rtjowc ft oo o

C0rt|T- rtCvCCN

r Zt u ft -U ft -4*

* -3 'rt O

*. - *n a. vt - cn cn :*j a. f,

uOJ-

f. if , </; *, ucl rt i- 4 if-

ift 0 0 VC I

rt m m m cn <

la, CN O CN CN 4 C W 4 C

r- a o 3

4- r< <t t Ul t a. < - M*P

*N*t _rj-^rv ,-. VI vf P-

if. P C P-_c m i

o o <

T J* vl -f I'

_T f. f 3 vt'

3 n y> *. o

-3 >3 _J N N 4 O

ir cto--

m m m m

Fl U. OC -rt rt

3a

a.

3 r .-- 9tn tr- V rt

4 O O

rt a -: 3 a n fttr. a c vt rv *r m

rt Jl .-* ft

vf CN vf C

if. m rt *. r- *- t P) V rt *- J o> CN f.

r* ir. If: hnp rt rt t- vC C rt , inccc

ft m 3 rt rtv- VL f *- P. ( - Fl CN vC

m *O & m u, t*. <

i It rt 3 n 4

i m a v.. in

.i.

k I- F. rt rt

i -rt ,j -rt .o n r-

y. #- a. tf .

m -. .r-

4* O N)-

F*m a. * <

at C 4 -

f'HTtr. 0 0Pntvux ^ _ vC a-tin i

) rt o ( * o r 4 o m vC 4 O N. ->

rt - * m '

IN 4 Irt rt I

-3

cn m

i r - i ft _? -1

4 *- CN ift r* m

r- o o

-atr- cn

rt

c. m m

f n /> ^ c -n iV* * t ? c m u

2"*- *p/i -0 -> 4 0

p- rt c.4-Nt_:cNC*'ctnmciNm

c> rt -) p 3

'

if ft r- rt ,

rt c c

c t uvo

t rt .M 3 rt] v

4. art Ot -* I

a O

O O - v

in 4 - -a.r o*

l rt J- JIP O

C U> O rt c

P*** itcvk

TO.

> at cd f

' O O 4 ?- p- cn

it Ft art)

on o

4 A. 4

4 J

CN) H3 - r-

t rt

O' m a* ot

vr x o a *(*%.

crr* p- rt rt

lu a a .

<n c_. 4 m __ m (

& -

14 O 3 -rt

'o> o_ m u> 4 c

-* * r o r i-N o * rt rtm i-l 4 r-

vc o 4 c r-

._3

. . . .

rt cn m c P- rt o oIrt O* )

X

u

O* X N 3 3 -* -* m r *rt

r- r-vt

p-? - 4 vf. CN

CN Oi C O 0 m ,

i*- f.-tft ir*iNotiO'rt4i

-rt |Mvp,jm jF Jl rt

A a lf> rt , U>_ afi U. If. F.

-*_!

fft mu ft c a. rt C

-trt ^

V

J-

c =3 Jl rt (VI PC jc p .VI <3 .N

m a v? 4 rt m <c rt m 4

U***^>

4 rt m vC rt - 3 a> Ot

cm 3

rcas-3 Rja

o 0 o

rtO*o

OvO * ft

o_t

4 P*D O

Jr* irtirt o mo m fcn I rt rt CN O O O

0 art4 ft ft9 AlA 0> i/t

cn r- rt m cn a. rt 0 rt m

3 4 rtOf- * 4 rt

ift o - mm

a* a a a a a aaaa

if fi m m m o mooo cn Ct O NO rt

(HP>*ft

4-* i>* <B S O 4>

MAPrt0* m in m cn cn m rt o * moi

* a a a a a a

Irtrt CN m rt |N CNOOO 4000

O CM -rt ft OB rt Oft)

IC4 <X> 0 IN OD CN

crmicm m -* m m 4 ic cc mco IN . rt 0 CN *a-.

9 OO ft -4rtrt)

MOWP- H.*-_P-

.

ccwm rt rtOt cNrt

m o

m m

!ft rt <*lO II rt 3(M SI (Nj^rtft04 0* F O) rtm 4 O v0 v0 ft

a a a a aaaa

CN rt O m CN CN CNOOO 4 O O

rNioii o ft ift rt 0043 p- ft m o* rvj rvi p~ p m o o 4cr 4 vc vc 0 cn r vt> cn 0 m at r- * 4 <c m cn ccr<fto

a a a a a a a a aaaa<-< o a, m 4 o m o* if <c 4 cn cn 4 o moo moo

lino** 01 o o - ft m~

v- P MQt vO OVO

act n* o cn irt rt m

OflTKM Ift aT m rt PI 4 rt rta> r> rt <-< p- m 4 r0icm

1 m m m vrtooo mooo

os m o 04 vjftom rt<N4rtit in ift m O* it -a it 4 m -rt m *p- m 0 0 m m ir or 4 4 mcr-m 4pnrtoi o ft *c o** aaaa aaaa aaaa aaaa aaaa

cn m cr 4 pbfifi r-vcmiN 4 p- m o mooo ft-<oo

r if, if -t co fi cn cn

O NF

CO c

O 4 rt rt ic m m m r-it> m 4 0. m rt cn

O (ft OCM O- > *** i- Pi CO

C*PP* IN IN m CD CN

t * aaaa

tF, to m CN m rt IN CN CN CT<vtCN rtCOO CN rt o

it P* P* 3 Ol 0 3 P- O N JiP 4 ^ 4 4

0rtcN mmm 4 rt4rt r- 4 rt .

CN C O C

ft ft ft P-

vC O

CNOO - m * m p* IN rt CN 4

0* m _o m 19 5trt04o o r* :m m

r* m cn 4 o# rt rtrtoo crm4rt 4CNom

i. *i m *I* O It C-

fma- INCC

c *r m a

m vc r fi C C C -ftC a CN rr m

B rt O Pi .4 3 ft c

C F. IN 4 IT. VC P- a

ft H MC cxp.m

tn r-

o. tr

r vt cn

j. 3 n 3

to m cn rt

c <f 0 o.

m c

.-1Nv*<V]vC&4vC M 143 P*

O

X

ft -ft -0 ,*- 3 * r NJ

. _ a -a -,

r.ft.n.

rt p* m m i- cn r*

j rt ft n r-

u ck, m c cn t

' iv vf: A v k, tf.

It 't 'ft ft <J> .!

r H cm m p- m

a vwft

Or* 0 ift

0 Ifi 4(_

"'Oft

c'

x o o .n m'p* p-

rt rt

c 4 m >i, m

cOr- *t

< p* Oi

CN Ov

_w

C - CN

Ft b. OD art rt

m O z 3 J- ft ft

ma c v_ inj-

m

* X X >! rt 1) N

f F. L vf c c m

r*- -* < 31 DO 'N

i/i Ft h r p k o

P* Ji X N! 3 O - -

cp- rt r- vt f-

F J 3 rt

- vt 4

I p- c m 01 wi . . tVlfift F. C. C

irccc vCFCmiTlNI-tCJ- '

IN "ft IN If rt C C

U 4

a cr

4 cn i/: o c- rfi cn

rt3NJftrt'rt

ca o c c m *rt

- w cv r

4 -3 m -rt

c

c* f 1 C u fi m r

4 P- art

P- rt Ci^(74^-OCF^-rtCNmCO o

IT-b'

- rt _ O ft '

I- CN

P. -T t

-a m

n rt 'N V| -rt ^ F V) -F 'J t ft n

n

*".*T Fl

O

c T ^ M (1.

rt

F^ p-c 3

vfc v_ v b. m vN SI *n c a>U~

4,

"3

r

X

a

.M c

a

rt

0 "N ft "N * 3 0 "1 * "M -rt r* 1

r O. vf m ^ IP t r- P p- a. C vC <T

I art 13.C'

V-.

raft c o 0F>c. aitra-cc rt a. in

OCNrtvj_,^CJ"

"" vl CN

*.atm rv.ar-rt

r- '. -ir* (. *r - tii r

rt rt,3 0 _rt -ft

c u. p a 0 o>

',rS ~Ct * p- p-__,

u:

0 0 a * O

n a in

c3 ft -rt iO 3

m

i0* *rt r+

xm

p-

-U

>rt

X

vO

rt

IN

-O * Irt *.-(!-.

a a. a tfi rt vc u- :

c f. a m a. o.

j rt

.

p- < F. ti

a. IX C OU O CNCC

_rt 3 ti rt rN iN

r* m a. 4 ^ m

4 m vc rt

c -*t

a

x 1- rt -3 'N

.A rt m 4

.5

5 U. CN

3

rtl

a

*

m '

-3

a. o <

v.

* vl

u C1 J

3* -?

I r" U. -* t_J '

O -J

"

: _.r-

o- m 1H f P ft ft V. . t ll CN O C- I

C JF

X J

O b-

fN 4. IN. if

H1

1

u -j

mtf

rt IN O O

|(n(-<"P*

JJIN 4 IN art CN moo

| BiNac'. ft * C*> 1*1 Fi IN IN 0

c * ft o

CV Ob ft ft

IN O O O

a. tc cn m

mooo

<?v0ft-rtOP" 4rtrt*4) H*fl

Z^^rt mmrtm r*cmcN r* f rt m

U*Urt IN > * 40 - c o mooo

mo ft vc vc m

| I/J arta-*

< ^ cn m *rt cn m

-a-. mcoo

a co a <_.

cc cc cc

_o n xi

ft cc cc

b ro a a co ca a

CO c o o

2co o

cc a c a c

o o

o c

o

HfflO CN 1*1 4 O rtrtOO rtOOO

a- c

c c c o o

FOOO ^ cC c r. c c

f-irtC-vCiCrtO 4 O 0 m.

1*0 t| ;t 0 0 * f ." 3 * O r* * vi -rt

"P if C If if p m o- m to cn rt rt F. 4

r>-i - h ifi p p. c iT. _r rt %r c

Jrt

u

3C

1 n p- 1: o o -rt d O P it 4 o * O o

fiOCdltuPrtJ 0 O p-ft u

r- 4 if-

U r, j rt a c. vt c o c vC c< N

'rtL"

pins^nran 0 ft -rt -NI rt 3 3 aP^NCft|0

VT 4 43 CN "> IP P V)5

1* - pj r cip- P> r* X c c o*

It ^ ** *I_5

v3_J

ma

'' CO > ft o1 1* m v a r*

if

-rt N 3 .N 3 N O -rt -rt) j_,* K< 0 4 ft 4 a< ti- r>

1 C IN J.1

U. rt O n 0 C t"' *~

ITrt -i 4 > 4

i/iC

X-rt t>

.JMrtrtOP- Tl ,

ft -"

_, t ft Tl 'O rt J ft N rt

f - 0 PI 0 u M . J CN rt rt rt 14? v. r- o

}*r-H ^o!lf*rt-1

vf CN Li. C rt " C

In" _

_.

I?t f ? MJIJIt*^f y, m

"OlSHlTi-v'

Jl JZ

Lu

tt c'-

P* rt) m 43

> at

fprti*iii

Ci m _. ;- Fi Ol rt lt> uj

t

3 (SOU

3 . K

n. cn rt q.1 i. rt -rt

._ -rt

C C C O 13

a t; c c_ r. c c

-a -j 1

_r >*-

:

UJ bJ <-. UJ b' bJ

I MA i,? F Jji-. = ^m-i.rrrmj

CnAt\ Cu'-i-PGi-: fZ,;-CuL;j:"'

'0 v,

'-A i .-v *

I Of 3

1

2

3

4

b

D

7

8

9

10

li

12

13

14

15

16

IMAGE

CON U Ei'.

CHAR

2i?zi4\z,si?l

2-*S*3^<e:l

1231324^2

1.3*3.1.

1 2U 1 2 12 1

1212

I 2 1 2U 3 . 1

I2I2iZ?_.'_-4_.1

121 2 1_-4ii

1 2 1 ^ 1 2 3 _

1214132-^^

12 1 2 1. _: 3 * :̂ ^ 1

123zl

2343?] 23^

2323?i?;Z-.?

FiLt. --jAi-.Pi.

C u N - R o '''

Ai-n rL,r un.

1_

*. 3 _

3_3

. .1 2 '

7. _? j 7 i

I ^ '4 j 4 i "? l

1 b 1 2 I _4 i 1

l.3a4j^i4j,Zl

1 2^24_.4_j?.i

1 2 1 2 3 2 3 i

J 2l2',2t

') 2 3 2 3 2 1

I 2 ? 2 ? _1

1 - 3 2 1 2 1

i. 2. 3 2 1

ij?i2j'>1

: 2 3 2 .3 _.

'i

W.iZcS F.jP 1 o v>;Vi

1

2

4

b

b

7

6

10

ll

12

li

14

lb

16

IMAGF

l.^-i^n 1

2^3.*?_

l23*3,?i

l23_3*2l

1 2 3 .. 3 v 3 <_ 3 2 3 i

U3-.3i3_1

123*3._1

) 23^hZi_^-.i

1..4-i4_ 3^/i?_

l_.34bZ?_?_l

123. 1

2i4a<ij2i'2l

U34321

l2.3H-323_

12354 iJj4.->?i

FILL. = *|P2.i""v,

CCNL/Fi-jScD riu-* AM? C-jTju-

CHAR CuN-f'L.W

1 I2343212-ij2j.

2 254_>2l232_?-

i 134^123231

123124321

b 1 2 3 1 3 2 1 2 ? 2 ? 2 1

o 2 3 A 3 1 2 3 2 3 2

7 123232l23^_

o123*32123-1

y 123.321

10 123-,''

11 1.3^.1

12 12321

13 12 3'. 3.

U

15 12404J1

16 12 343.

! 2 3 4 3 / 3 _.3 4 i

-23-,323</3A\

12^3232^2^

1 2 ^ - 3 ; ^ 2 ? 2 '.

I 2 3^i 32 "2^21

1 2 3.3 2 ? 2 3 2 !

1 2 3 H 3 _

? 2 3 2 1 2

1 2K323232"

1 .

^ 2 3 3_ I

3 i323i?. '2

1 2 3 t.323,323x

i 2 3 2 3 t 3 1 1

1 2 3 _ 3 '_ 3 _ 3 2

1^3-,32.3-.3il

12 *2 3-*3 23*!

' l 3 . 3 ^ 3 i 3 ^ 'i 2 I

ii.'1-."! r o'-

Cu.'.-CoLu -'

12 3 2 3 2 1

1. 2 3 2 3 2 1

12 3*32

123232'

\_3232l

l2?232l

'li'li")

I 2 I 2 1 2 I

I_

1 2 1

1.1.1

I 2 1 2 1Z-l-1

1.1-1

I 2 1 2 1

2 1 2 '

I- Lh>n'

IrtAoF FILt =

CuNLFi.Sc-P1'u1

ChArvCuri-fZ.'

st.-1-;.

5-*?

v.( Z

-cn

I 12 12121

2 1 2 1 2 1 2 1

3 12WUI

4 1212121 2

b 1 2 1 2 1 2 3 1

6 121 2I2I

7 I2I2IZ

b 1 2I 2l_3_

* 121 2

10 1212

11 1212

12 12121

13 121

14 I2I2I

lb U12J

1 o 1 2 1 2 1

IMAGE FILL = 5&I'-

CON-jFimScD *uW A.-

ChAn CuW-Pu"

- I VJ

Cl J .0'

1 1212 12 3 21

2 12 13 21

3 I2IJ?.

4 1 2 1 2 1 2 1 2 3 .

1

5 1 2 1 2 1 2 1 .? i

6 121.1.U3J

7 l.l^l.l.J.l

0 1 2 1. 2 132i

9 1.1.UUV/j?i

10 l*l*)2 3<,3.2i

11 I2l2l23<.c;-2i

12 1212l2.3 43._l

13 121234^321

14 1 2 1 2 1 2 3 4 ?

15 12.1_.U34?. J

16 12U 12343.1

IVAGE FILt =_>Jtj}.Pi.i'-i_

Cjr;DEi.ScP ku''" A;.D CuLuZj

CrtAR CLM-Ro^

1 12121*3*

2 1 2 1 2 1 2 3 * 1

3 121312 3 2

4 1 2 1 2 1 2 3 2 1

5 li.l_.-23i

b 1 * 1 2 1 2 3<* ? i

7 I2I2I 23i2

_ ia-.i.-_i

9 U31?i?.?j

10 1212 3 2 1

11 12121 23^ 1

I2 1 2 1 2 I 2 3 * 1

13 1 2 1 2 1 2 3 * 1

14 1_l2l2321

lb 1 2 1 2 1 2 3 2

lo 1 2. 1 2 3_

'-11 ? 4 3 _ ^ 1

12 .

'_343/i

1 2 1 3 4 2 1

2 3 * 3 _ ? i12 3 -. 3 1 3 i

_ 2 3 -r 3 _ 2

12 3 4 3 4 3/.

- 2 3 1 3 n 3 2 11 ./

1*1*12 1

1*1*1

1212

1*121

! 2 1 2 1

1.1.1

1 2 1 2 1

12 3 212 1

\ j- !-'c. S F j L 1

I'O^-Cur.y ..

1 *3^3^Sa^ /

U34?

1 2 4 j 4 j ? 1

12 3 -. 3 1 S , 3

c

!. 4 .4 0 ? i

1 2 3 * 3 ;. 3 * 1

1 * 124 j- 3 u

1 2 5 3 4 b -1 * 1

i ,.3*3 4 ^-Zj.

1 * 3 2 3 <_ 5 l. 1

; 2 .1 2 j 4 3 '1 Z

12 3 2 3 4 3*'

12324j?l

l 2 3 2 3 t .3 2 '<

'

2 3 2 3 _ . -, 3 * 1

"

2 3 2 3 :, 3 * 1

CWnS-r.f,,s

.n,ZLS Fu^

Cu^-CuLZ.-

12-123/1

-232!

T 2 ? 2 1

U 3 * 3 *1

: * 3 2 3 * 1

12 3 2 3 2 1

j 2 3 2 3 2 3 * 1

_ _ 1 * 3 2 3 2 1'

1*3*1

1 * 3 2 3 -. 3 2 1

' * 3 * 1

t 3 I 2 3 2 1

1*1*3*1

'

2I 23*1

.*l*3*l

*1*3.1

lo v.HZ

IaAgF FILt = iAi.Po. j.' t

CONLEuStO H.ji-

;, I-, C'ur.^f-

CnArt CuN-kwl-.

5"

<7 ?of?

,sFur !c LH:rj^

CliLj'-Z

4

5

o

7

6

9

10

11

12

13

14

15

16

Ii'iAG

CON'U

ChAr.

U

12

12

12

U

1*

1*

12

12

12

12

12

12

12

U

13

E FI

Fi'.Si-

Cb

Ul

121

Ul

Ul

Ul

132

312

123

123

123

123

Ul

Ul

121

123

121

I.S.

D P

N-R

*3<.3Z U32 3*l22 3-.3..1 3*3*3*1

2342_ 1 23*_.:1

2.3<_3Z j 2 3 * 3 * 1

23*3. J2323*l

l2ia 2 3 2 3 2 1

12ja;-V 2?.232l

42i 3 2 3 2 3 * 1 *

4 321 1 32321

-.32 2 3 2 1 2

431 ?. 2 3 2 3 * l

23*321 *

2 3 2 3 2 1

23*2i *:--J?l

2 3,3*1 12 3 2 3 2 1

4321 '' 2 3 2 3 2 1

23432 ;2323_;1

= if,P/.j"i,

\jV A,.D f jl;u'\'. VJiLil't-K rZ

Uui (Z--CuLu.ii

1 12321

2 U343212

3 123*32 1

4 234j2i

5 1234 321

6 1*343*1

7 234321

o 232321

9 2 4 3 2U 3O

10 l243U3s3_l

11 123<*3*U3~3*

U 2 4 3 i 2 _4 _, ? i

13 12321*3.21

1 4 U3i2 34*

15 U32U2Z..1

16 U432l?_-2.

IkAGE FILt. =

CLiNCEi-jSi-n kuV-'

ChAnCuM-Pc>.'

id.n r.\

lo U'<ZZ'

12 3*1*1

I * 3 2 1 /. 1

12 3 2 1*1

12 3 2 1 * 1

'. 2 3 2 1 _ 1

i23*l*l

i.3-1*1

1 23*1*1

.1 2 . 2 1

1*3*1

2j?i2

'1321

1*321

Z?_

->_-2A

'2321

vZl'i-S C*JR lo v.HM'Zr.

Cu .-C'jLJ'K

1 U3b323..3,.._ >34j?l?-2i

2 U?b3*3^Z.l ij'7l?..i

3 U32343-.2 i 23t3*U3fc'i

4 12 32 3 2 3^3*1'' 3 4 j / i 2 j 1

5 234j2j4^/0;2 '. i 2 1 / i 2 i

o 123-4^*3.,__.?rt3*L j 4 3 ? 1 2 3 1

7 1233323^3-3*1 i24-Z23?-

e 2^323*3-3. i 2 3 * 3 * 1 * 3 . 1

y 12323* 1*3.; 3* 12323*1

.01.3.3.U3:? ! 2 3 2 3 _

/

11 123232123^3* ''3?J'?i

U 2j23231*3-,3*i ?3?3?i

13 1* 32 32 12 3-3 j, 2:52.521

14 l*4j?3?j?.'42 '>32-io1

15 1.3^.3.3.. 1'2 3 2 3 * 3 *

lo 1232321*3^22323421

IMAGF FU..,. = .r'J,rr.ri-T,r.

CUf-iuFi.StnHui': Ar.i?

ruT,L,.

CMAK Cur'-Ru*

5 ic rposy-Sec-no^s:i of.

* Fu^ 1 o

1 24M3*1

2 2*21

3 2 j 1 * 3 *

4 12 34 3 2

5 1-1.121

6 1*12

7 12121321

o 121212 34 2

9 121*3/

10 1 2 1 2 3 *

ll 121*3*?

12 UU U 3 2

13 2321

14 2 j l 2 3 4 ?

lb 2.2j"_

2i2j_._?j?

FILt, =_.!.:_

Id

IMAGE _..

CuNUFijStP Rl.

CriA,

1

2

3

4

5

o

7

b

9

10

11

12

13

14

lb

lo

IMAGE

Ci-l-:-Pi

U .3 * * -i 3 * I

2-.3*1

1 *343*1

1 2 4 j ? x

1 24_,?i

2*3*

.1 * 4 j 2 .

14 3*1

2 3 '1.. 4 _ 1

1 j. 4 a 4 *

2<Zi

2 .. 4 :> 4 * 3 * 1

U'Z?_

1 j4j?i

FlLi_ -_jA;

CundfnocD puw a;

CHAh CuiJ-RijV;

j.'VJ

C.1 ul

'jf.,j

i lj4i?j.?j?_

2 2h.32U32

3 1- 1 * 3 2 1

4 U 3 i 3 2

b 1 j 2 1 2 .

o 3.2 j?

7 U21232

b 131*3*

9 12 3*

10 U3s2

11 12 3

12 12 3

13 12 31

14 1.3.2

lb U3_

16 12 3i

Cu/'-CuLu'-'..

' /.3*3*1

U3_323

ni?s7

1 21

.;*34i

1212

1 2 5 3 4 -j i l

1 2 324*1

1*1

731.i1 *32321

13 2 3 2i

1*3*1*1

'>l7i

1 321231

3*3.

VfUZtSf'jl'

Id C^rt1'*V~

J. i'.^

Cuy-CuLu"'i-

1*3*1

1*3432**1

1*3*3*3*1

1*32321

I 2 3 4 3 2 3 * 1

12 3 2 3 21

L/34323*1

1 / 3 * 3 -i 3 *

U 3 4 3 2 3 -* 3_

1. /. 3 * 3 4 3 2 1

i. 2 3 * 1 * 3 * 1

1 2 3 4 3 * 3 /* 3 c

1*3^3*3 'j

1 _234_-2

1 j 4 j '2 _i 4 j 1

**jjijt_SFu(

Cu.'i-CuLiJ'

l_2_2l

2j2ji

.1 2 3 2 3 2

2 j 2 3 2

3*3

Zi2->->

i _.:. 2 3 2 1

'

. j 2 J ?.

Ul.l

1*12 1

I * i * 1

I * 1 2 1

UJ 2l

1 2 1 2 1

1*1*1

U U 1

lb u H ._,

r-

b-)D o-crl

IMAGE FiLc. =_,,. U

Cjndf: i.StP Z'-' V .P TuiLvj

ChAh Cul'-Zi-

---_ -------

1 12121*1

2 U 1 2 i * 1

3 U 1 2 1.

1

^ 12121

b 1.1-1 *1

c Ul*l*i

7 U 1 2U 1

0 1*1*1

Ci 121*

10 121*

li 1*12

12 UU1

13 Ul

14 1 2 1, '. 1

lb uu

lo 1*1*1

Ii-iAu-F FILt = : P c . .1: o

CiZl-Fj.,.s_.n^ ..- v .^ ZT,j

CriAhcgi-.'-p_.i-

----. -_.--.-__

1 121 23*1

2 l_-?_

3 U2i

'i 1..1

b 1..21

o 1*3*1

7 12 3 2 1

c U2i

9 UU 3. ?

10 1.3.3.?

11 12 12 1 * 3 -i 1

1 2 U U .3* 3 2

13 1*432

14 U123i2

lb 1212 3*3

io UUU3-.3. : 1

IwAGE Fj.Lt - oA. i f J . *'

>u

'-?-.3t32

U3.

?-3-_2

L.>4^1

3J-.3-* 1

1-.3-.2

3-i3

li 4 j 1 j "Jx

1.1.1

I. 2U

1..U

12 1 2 1

-

1 2U 1

1*1*

'

I _.> ..j1

:-C..L,J'

U 'i j

US-.

1 24 J

,_

4 3

Z34

i *4i

i*4.

_ 2 3-i

I 23*

1 *32

U3*

L232

1*3*

1*3*

\ 2 3 2

'2 32

4*

*

3i

4 2 1

4.

54?

4 b 4 o 1

345*1

34 31

3 ..3i

4.?

3-,3*1

4j?

3rt?l

CUl-DEjS_.0 Pu'i

CrlAH CuN-Ro'-i

(J I j u ' i-Hi-S FuP lo

u '-CliI. u <'-..

1 1 * U: 1 2 3 2

* 1 2 12 1* 3 i

3 UU3*

4 U 1 2 3 2

b 1_1.3.

0 1.1.3

7 1.1.3

a U 1 * 3 *

9 121 *3_

10 hl.3

11 U3i

i. 2 1*1*3.

13 1 *1*3

14 1-3

lb 1*1*3

1. 1*1*3

ii?i2_

U.321

1.321

3*321

13232.

1*321

lZ>i

1*321

12321

1*3*1

l_3.1

123*1

1.1.3.1

1 2 3 * 1

I*l*3*1

1*321

>3)C ic? 3

IMAGE FILL =-.A,-iPo.ifG

CONDEiiStD ROW AnD CuLuM.\i vALUES FOR lo CKmRaC'I'KkS

CriAh CON-ROW CON-COLUMim

1 U1212343 1232321

2 1212342 1232321

3 12U342 1232321

4 123* 1232321

5 12123431 1232321

6 1212342 1232321

7 UUU342 1232321

6 1212342 1232321

9 12123431 1321

10 1212342 12321

11 123431 1232321

12 1212342 1232321

13 12123*2 132321

14 UU3*2 1232321

15 123431 U3232.1

16 12342 132321

IMAGE FlLc. =_.Ai'iP7.IHG

CONDENSED ROW A^D CuLUMiM VALUES FOR lo CHaRhC'IEkS

CHAh CON-ROW CCN-CGLUMft

1 2321 1*32121

2 234321 1232121

3 U3t3*l 1232121

4 232i 1232121

5 232. 1232121

b 234321 1232121

7 234321 1232121

S 2321 1232121

9 2312342 321

10 243-2j42 321

U 2431234* 2321

12 24312343 321

13 23212342 2321

14 312342 2321

15 2312342 2321

lo 2431232 232i

IMAGE FILE = oAi.Pd.iMta

CONDENSED ROW A.^lD CULuMi. VALUES FOR lo CHaRmC'iE^s

CHAR CON-ROW CuN-CuLUMn

1 23234342 23212321

2 2323431 13212321

3 2323432 3212321

4 1232342 2321231

5 243234342 23212^21

6 13234342 132131

7 232342 2321231

b 2432342 1321231

9 232321232 132321

10 232321234* 13232

11 2321232 232321

12 232321232 232-.1

13 1321231 23232

14 23212342 3231

15 12321232 23232

lo 23232123? 23231