Lecture 1 - Faculty of Engineering and Applied Science ...charlesr/9881/lecture1.pdf · Graduate...

ENG 8801/9881 - Special Topics in Computer Engineering: Pattern Recognition

Memorial University of Newfoundland

Pattern RecognitionLecture 1

May 2, 2006

http://www.engr.mun.ca/~charlesr

Office Hours: Tuesdays & Thursdays 8:30 - 9:30 PM

EN-3026

ENG 8801/9881 - Special Topics in Computer Engineering: Pattern Recognition 2

Course Outline

This course covers topics related to the recognition of unknown patterns based on

distinguishing properties. These topics include the treatment of patterns as random vectors,

distance measures, MED classifier, ortho-normal whitening, MICD classification, MAP

classification, Bayes’ classifier, bounds on performance, parameter estimation and density

estimation, linear discriminants, hierarchical clustering, iterative clustering, feature selection

and extraction.

Background

Linear Algebra, Calculus, Probability and Statistics, Programming

Instructor

Dr. Charles Robertson 368-9984 (w) 726-2892 (h) [email protected]

Office hours: EN-3026, Tuesdays and Thursdays from 8:30 - 9:30 PM, or by appointment.

Lectures

Tuesdays and Thursday evenings, 7:00 PM - 8:15 PM EN-1000

Textbook

Pattern Classification, 2nd Edition by Duda, Hart, and Stork; published by John Wiley, 2001.


8801 9881

Midterm (1.5 hrs) 25% 20%

Final (3 hrs) 50% 25%

Assignments (5) 25% 15%

Project N/A 40%

Assessment

All students in both 8801 and 9881 will have one midterm test and one final exam. There will be

five assignments, which are mandatory for all students. You must submit all assignments to pass

the course. Students in the graduate course will also have to complete a project.


Graduate Student Project

Students in 9881 must pick a problem where some area of pattern recognition can be used to

solve it. You must research the topic, create an application, give a short presentation, and submit

written reports. The report topic should be a couple of sentences on the nature of the problem.

Abstracts should be several paragraphs and include a list of references. The presentation will be

about 10 minutes total, including a couple of minutes for questions. The final report should be

30-40 pages long, not including appendices.

Implementation of pattern recognition algorithm(s) is necessary. Use MATLAB, C++, or Java as

your programming language, or check with me for suitability of other languages.

Sample topics and problem areas:

- edge detection in images, texture classifier, character recognition, document layout

processing, facial images, medical images, industrial inspection, affect of noise on images,

finding objects in SAR images, sketch recognition

- audio/video signal processing, signal processing (biomedical, mechanical, geological, etc.),

biometrics, music, word-analysis, speech processing, speaker recognition, meteorology,

security

Remember that your project may not be used for credit in other courses or for your thesis.


“Car Detection in Low Resolution Aerial Image”

-1

0

-1

-2

-1

0

0

0 0 0 10

2 -2 0

0 1 2 -1

-1

-1 0 0 0 1

0 1 -1 0 1

1 2 -2 -1 0

1 2 -1 0 1

0 0 0 1 2-1

1

0

0

“Automatic Processing of Information on Cheques”

TUPIN et al.: DETECTION OF LINEAR FEATURES IN SAR IMAGES 451

(a)

(b)

Fig. 21. Road detection process on a RADARSAT image. (a) Original RADARSAT image c Canadian Space Agency (available on the CD-ROM RadarsatInternational). The pixel spacing is 12.5 m with three looks. This is an image of the Amsterdam city (The Netherlands) with many roads and channels.(b) Result of road detection superimposed on the RADARSAT image. The main axes of the road network and hydrological linear structures are detected.

“Detection of Linear Features in SAR Images”


336 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 50, NO. 3, MARCH 2003

(a) (b) (c)

(d) (e) (f)

Fig. 1. Different kinds of anomalies in the ESDB database. Upper panel represents the ECG with the originally annotated beat classification. Lower panel showsthe original heart rate and the modified one indicating which action has been carried out. (a) FP deletion. (b) FN insertion. (c) Supraventricular ectopic beat moving.(d) Ventricular ectopic beat moving. (e) Deletion and insertion in one-step double correction. (f) Multiple insertion of evenly spaced beats.

HP or heart rate HR or low-pass filtered event series (LPFES).

The continuous time generalization of the IPFM model can be

written as

(2)

where is a continuous function that solves the model equa-

tion and whose values at

are the th beat occurrence time. can be seen as

the instantaneous heart rate . is the mean RR interval in

the analyzed period and represents the dynamic part

of the instantaneous heart rate, which is zero-mean. The dy-

namic part is usually small compared with the mean heart rate

. The first beat is considered to occur at and

is assumed to be causal, hence, if . An-

other important restriction imposed on the signal is that it

is a band-limited signal with negligible power spectral density

over 0.4 Hz.

The heart timing signal is also introduced in [8] as the

integral of the modulating signal

(3)

and, from (2), we obtain

(4)

This signal defined in (3), and observable at according to

(4), is then used to estimate the PSD of the HRV, since

can be estimated as .

An ectopic electrical impulse may cause the premature reset-

ting of the integration process on the SA node, giving a phase-

shifted series of normal beats following the ectopic beat. If the

ectopic beat is the th, it can be interpreted that the beats pre-

ceding the ectopic beat are given at with integer and

, and the beats subsequent to the ectopic beat are given

at with integer and . The prema-

ture resetting of the integration occurs at and

is an unknown real quantity corresponding to the value

reached for the integral at the resetting time . It is

important not to confuse the unknown instant with

the known (estimate) ectopic beat occurrence time, because the

propagation mechanisms of the ectopic impulse may be some-

what different to the normal beat and then, the instant of the

SA node reset will not be the same as the one

associated with the ectopic beat position, typically from a QRS

detector. The time occurrence of the ectopic beat has no relation

to the SA node and it will not be used by the proposed method.

Fig. 2 shows the signals involved in this IPFM extension. The

dashed curve represents the function given by (2). The con-

tinuous curve is the same function but reset when it reaches the

corresponding threshold. The normal threshold is one but the

ectopic beat prematurely resets the integral process at . The

subsequent beats will correspond to values of coming from

integers to which have been added an unknown constant mag-

nitude .

“Analysis of Heart Rate Variability in the Presence of Ectopic Beats Using the Heart Timing Signal”

1 2

3 4

100

80

60

40

20

0

false acceptance

105-5threshold

0

false rejection

% errors

“Acoustic-Labial Speaker Verification”

0

5

10

15

20

10000100010010Size of database

Ave

rag

e n

um

be

r o

f n

ote

s

req

uire

d f

or

un

iqu

e id

en

tifica

tio

n

Figure 4. Number of notes for unique tune retrievalin databases of different sizes. Lines correspond,from bottom to top, to the matching regimeslisted in Figure 1. Figure 5. Display from the tune retrieval system

Based on these dimensions, we have examined exactmatching of:

• interval and rhythm;• contour and rhythm;• interval regardless of rhythm;• contour regardless of rhythm;

and approximate matching of:• interval and rhythm;• contour and rhythm.

For each matching regime we imagine a user singing thebeginning of a melody, comprising a certain number ofnotes, and asking for it to be identified in the database. If itis in the database, how many other melodies that begin thisway might be expected? We examined this question bytaking every sequence of n notes that appear as thebeginning of a melody in the database, and, for eachsequence, counting the average number cn of “collisions”—

that is, other melodies that match. In order to controlcomputation time using dynamic programming heuristics tospeed the match (Waibel and Yegnanarayana, 1983),fragmentation and consolidation were not considered in thisanalysis.

Figure 3 shows the expected number of collisionsplotted against n, for each of the matching regimes. Thenumber of notes required to reduce the collisions to anygiven level increases monotonically as the matching criteriaweaken. All exact-matching regimes require fewer notes fora given level of identification than all approximate-matching regimes. Within each group the number of notesdecreases as more information is used: if rhythm isincluded, and if interval is used instead of contour. Forexample, for exact matching with rhythm included, ifcontour is used instead of interval an extra note is needed toreduce the average number of items retrieved to one.Removing rhythmic information increases the number ofnotes needed for unique identification by about two ifinterval is used and about five if contour is used. A similarpicture emerges for approximate matching except that thenote sequences required are considerably longer.

An important consideration is how the sequence lengthsrequired for retrieval scale with the size of the database.Figure 4 shows the results, averaged over 50 runs, obtainedby testing smaller databases extracted at random from thecollection. The number of notes required for retrieval seems

to scale logarithmically with database size, although all thelines do bend upwards towards the right, indicating thatlonger sequences may be required for very large databases.

A system for tune retrievalWe have developed a system, based on the melodytranscription program described above, for retrieving tunesfrom the combined Essen and Digital Tradition folksongdatabase. The user starts singing on any note, and the inputis notated in the key that yields the fewest accidentals.Transcription operates in adaptive mode, adjusting to theuser’s gradually changing tuning.

The user is able to retrieve folk tunes using an exactmatch of pitch, pitch and rhythm, melodic contour orcontour and rhythm, or an approximate match of pitch andrhythm or contour and rhythm. All searching incorporatesfragmentation and consolidation, as described by Mongeauand Sankoff (1990), and all retrievals are ranked—exactretrieval simply means that only tunes that match with themaximum possible score are retrieved. Our dynamicprogramming match algorithm is a minimization technique,with the perfect score being zero. In order to provide amore intuitive score for users, the dynamic programmingscore of each song is subtracted from 1000, with a lowerlimit of 0, so scores can range from 0 to 1000. Songs areranked by score; songs with equal scores are listedalphabetically by title.

Figure 5 shows the tune retrieval screen following asearch. The names and corresponding scores of retrievedmelodies are displayed in a text window and the tune of thebest match is displayed in a melody window. The user mayselect other melodies from the list for display. In the Figure,the user has selected Three Blind Mice, following anapproximate search on pitch and rhythm. The large numberof songs retrieved is not surprising in light of Figure 3,which shows that, for five notes, approximate matching ofpitch and rhythm can be expected to yield thousands ofcollisions. The number of collisions can be controlled,however, using a variable retrieval threshold.

While the system currently matches all tunes from thebeginning, we plan to extend the matching algorithm toallow retrieval based on themes which may occur anywhere

“Towards the Digital Music Library:

Tune Retrieval from Acoustic Input”


[1] Automatic processing of information on cheques. Systems, Man and Cybernetics, 1995.’Intelligent Sys-tems for the 21st Century’., IEEE International Conference on, 3, 1995.

[2] Towards the digital music library: tune retrieval from acoustic input. Proceedings of the first ACMinternational conference on Digital libraries, pages 11–18, 1996.

[3] Acoustic-labial speaker verification. Pattern Recognition Letters, 18(9):853–858, 1997.

[4] Detection of linear features in SAR images: application to road network extraction. Geoscience andRemote Sensing, IEEE Transactions on, 36(2):434–453, 1998.

[5] Car detection in low resolution aerial image. Computer Vision, 2001. ICCV 2001. Proceedings. EighthIEEE International Conference on, 1:710–717, 2001.

[6] Analysis of heart rate variability in the presence of ectopic beats using the heart timing signal. BiomedicalEngineering, IEEE Transactions on, 50(3):334–343, 2003.

References


Important Dates

May 9 Report topic due (2-3 sentences)

May 18 Assignment 1 due Report abstracts due (1 page)

June 1 Assignment 2 due

June 15 Assignment 3 due

June 15 No Class Progress report due (3 pages)

June 27 Midterm Exam

July 6 Assignment 4 due

July 20 Assignment 5 due

July 25 Reports due (30-40 pages)


Academic Dishonesty

The university takes a strong stand against academic dishonesty, which includes cheating,

impersonation, plagiarism, theft of examination papers and other materials, use and/or

distribution of stolen material, submitting false information, and submitting the same material

for more than one course.

Any assignments or exams that are copied or otherwise plagiarized will not be credited towards

your final mark. It is extremely important that any material you include in your project reports

are correctly attributed to their original sources.


An Introduction to Pattern Recognition

What is Pattern Recognition?

“The act of taking raw data and making an action

based on the ‘category’ of the pattern.”

Two key approaches

Classification - we already know the classes

Clustering - we don’t know the kinds of classes


An example of

Industrial Inspection


FIGURE 1.1. The objects to be classified are first sensed by a transducer (camera),whose signals are preprocessed. Next the features are extracted and finally the clas-sification is emitted, here either “salmon” or “sea bass.” Although the information flowis often chosen to be from the source to the classifier, some systems employ informationflow in which earlier levels of processing can be altered based on the tentative or pre-liminary response in later levels (gray arrows). Yet others combine two or more stagesinto a unified step, such as simultaneous segmentation and feature extraction. From:Richard O. Duda, Peter E. Hart, and David G. Stork, Pattern Classification. Copyrightc! 2001 by John Wiley & Sons, Inc.

12

Richard O. Duda, Peter E. Hart, and David G. Stork, Pattern Classification. Copyright 2001 by John Wiley & Sons, Inc.


Model Design

We need to develop models of the fish that can be used

to distinguish them.

Training samples can be used to find distributions of

features.


salmon sea bass

length

count

l*

0

2

4

68

10

12

1618

20

22

5 10 2015 25

FIGURE 1.2. Histograms for the length feature for the two categories. No single thresh-old value of the length will serve to unambiguously discriminate between the two cat-egories; using length alone, we will have some errors. The value marked l! will lead tothe smallest number of errors, on average. From: Richard O. Duda, Peter E. Hart, andDavid G. Stork, Pattern Classification. Copyright c" 2001 by John Wiley & Sons, Inc.

14



2 4 6 8 100

2

4

6

8

10

12

14

lightness

count

x*

salmon sea bass

FIGURE 1.3. Histograms for the lightness feature for the two categories. No singlethreshold value x! (decision boundary) will serve to unambiguously discriminate be-tween the two categories; using lightness alone, we will have some errors. The value x!

marked will lead to the smallest number of errors, on average. From: Richard O. Duda,Peter E. Hart, and David G. Stork, Pattern Classification. Copyright c" 2001 by JohnWiley & Sons, Inc.

15



2 4 6 8 1014

15

16

17

18

19

20

21

22width

lightness

salmon sea bass

FIGURE 1.4. The two features of lightness and width for sea bass and salmon. The darkline could serve as a decision boundary of our classifier. Overall classification error onthe data shown is lower than if we use only one feature as in Fig. 1.3, but there willstill be some errors. From: Richard O. Duda, Peter E. Hart, and David G. Stork, PatternClassification. Copyright c! 2001 by John Wiley & Sons, Inc.

16



?

2 4 6 8 1014

15

16

17

18

19

20

21

22width

lightness

salmon sea bass

FIGURE 1.5. Overly complex models for the fish will lead to decision boundaries thatare complicated. While such a decision may lead to perfect classification of our trainingsamples, it would lead to poor performance on future patterns. The novel test pointmarked ? is evidently most likely a salmon, whereas the complex decision boundaryshown leads it to be classified as a sea bass. From: Richard O. Duda, Peter E. Hart, andDavid G. Stork, Pattern Classification. Copyright c! 2001 by John Wiley & Sons, Inc.

17



2 4 6 8 1014

15

16

17

18

19

20

21

22width

lightness

salmon sea bass

FIGURE 1.6. The decision boundary shown might represent the optimal tradeoff be-tween performance on the training set and simplicity of classifier, thereby giving thehighest accuracy on new patterns. From: Richard O. Duda, Peter E. Hart, and David G.Stork, Pattern Classification. Copyright c! 2001 by John Wiley & Sons, Inc.

18



Some Lessons Learnt

Training samples help define the models

Some features are more discriminating than others

Multiple features can be used for better models

Some decisions have higher costs

General models perform better than overly complex

models


An Introduction to Pattern Recognition

General Definition

Given some measurements of an unknown pattern, find

a useful representation and then assign it to a class or

category based on distinguishing properties.

20

Steps

DATA Sensor LabelsFeature

Extraction

Measurements

Classification

Features

Knowledge

Date post:	16-May-2018
Category:	Documents
Upload:	hoangmien
View:	213 times
Download:	0 times

Lecture 1 - Faculty of Engineering and Applied Science ...charlesr/9881/lecture1.pdf · Graduate...

Documents