Date post: | 20-Jan-2023 |
Category: |
Documents |
Upload: | khangminh22 |
View: | 0 times |
Download: | 0 times |
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 [email protected].
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.
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.
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
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
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
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
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
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
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