Dr' Claude C' Chibelushi

1
Fac. of Comp., Eng. Tech. Staffordshire
University
Image Processing, Computer Vision, and Pattern
Recognition
Statistical Pattern RecognitionPart II
Classical Model
Dr. Claude C. Chibelushi
2
Outline

• Introduction
• Classical Pattern Recognition Model
• Feature Extraction
• Classification
• Applications
• Optical Character Recognition
• Others
• Summary

3
Introduction

• Pattern recognition often consists of sequence of
  processes
• often configured according to classical pattern
  recognition model

4
Classical Recognition Model
Simplified block diagram
Facial image
Example
5
Classical Recognition Model

• Performance issues
• All stages of recognition pipeline, and their
  connection, affect performance
• typical performance measures recognition
  accuracy, speed, storage requirements
• optimisation of components/connections often
  required
• careful selection / design / implementation of
• data capture equipment / environment
• processing techniques

6
Feature Extraction

• Aim to capture discriminant characteristics of
  pattern
• Extracts pattern descriptors from raw data
• descriptors should contain information most
  relevant to recognition task
• descriptors may be numerical (quantitative ) or
  linguistic
• group of numerical descriptors often known as
  feature vector

7
Feature Extraction

• Common features for computer vision
• Shape descriptors
• external (e.g. boundary) internal (e.g. holes)
• Surface descriptors
• texture, brightness, colour, ...
• Spatial configuration descriptors
• arrangement of basic elements
• Temporal configuration descriptors
• deformation or motion of basic elements

8
Feature Extraction

• Example
• Pattern recognition application gender detection
• Classes male, female

9
Feature Extraction

• Example (ctd.)
• Chosen features height, silhouette area

10
Feature Extraction

• Example (ctd.) pseudo code
• frontEnd(image)
• // foreground-background image segmentation
  (e.g. thresholding possibly after noise removal)
• prpImage preprocess(image)
• // calculate height and width of image segments
• features FeatExtr(prpImage)
• return features

11
Feature Extraction
Graphical representation of feature
distribution Example (ctd.) data set of 5 male
and 5 female subjects
12
Feature Extraction

• Graphical representation of feature distribution
• Example (ctd.) Feature plot 2D feature space

13
Classification

• Aim to identify class (category ) to which
  unknown pattern belongs
• Wide variety of classifiers
• Classifier selection is problem-dependent
• use simple classifier if effective

14
Classification

• Some classifiers
• Minimum-distance classifier
• classification based on distance from
  class-prototype (e.g. average) pattern
• closest prototype determines class
• k-nearest neighbour classifier
• classification based on distance from class
  patterns (or clusters)
• closest k patterns (or clusters) determine class

15
Classification

• Some classifiers
• Bayesian classifier
• classification based on probability of belonging
  to class
• most likely class
• Artificial neural network classifier
• classification based on neuron activations (shown
  to relate to class probability)
• most likely class

16
Classification
Minimum-distance classifier
2D feature space
17
Classification

• Minimum-distance classifier pseudo code
• minDistClassifier(unknownFeatVect, prototypes)
• minDist MAX
• class UNKNOWN
• // assign unknown sample to class of nearest
  prototype
• for each protoVect in prototypes
• dist distance(unknownFeatVect, protoVect)
• if (dist lt minDist)
• minDist dist
• class class of protoVect
• return class

18
Classification

• k-nearest neighbour classifier

2D feature space
19
Classification

• k-nearest neighbour classifier pseudo code
• kNNClassifier(unknownFeatVect, dataSamples, k)
• size k
• initialise(minDist, minDistClasses, size)
• class UNKNOWN
• // find k samples nearest to unknown sample
• for each sampleVect in dataSamples
• dist distance(unknownFeatVect, sampleVect)
• updateKNearestDist(dist, minDist,
  minDistClasses)
• class majorityClass(minDistClasses)
• return class

20
Classification

• Some distance metrics
• (for distance-based classifiers)
• Measure similarity between unknown pattern and
  prototype pattern
• based on differences between corresponding
  features in both patterns, e.g.
• Euclidean distance sum of squares of differences
• City-block (Manhattan or taxi-cab) distance sum
  of absolute values of differences

21
Classification

• Decision boundary for
• minimum-distance classifier

2D feature space
22
Classification

• Limitations of minimum-distance classifier
• Prone to misclassification for
• high feature correlation
• problems requiring non-linear decision boundary,
  e.g.
• curved decision boundary
• data with subclasses (i.e. clusters)
• intricate decision boundary

23
Classification
2D feature space
Minimum-distance classifier feature correlation
24
Classification
2D feature space
Minimum-distance classifier curved decision
boundary
25
Classification
2D feature space
Minimum-distance classifier distinct subclasses
26
Classification
2D feature space
Minimum-distance classifier complex decision
boundary
27
Classification

• Bayesian classifier
• Bayes rule P(CF) ( P(FC) x P(C) ) / P(F)
• P(CF) a posteriori probability that observed
  feature F is from pattern that belongs to class C
• P(FC) conditional probability of observing
  feature F given class C
• P(C) a priori probability that randomly-drawn
  feature is from pattern that belongs to class C
• P(F) total probability of observing feature F
• for mutually exclusive classes P(F) ?c (P(FC)
  x P(C))
• P(F) is class independent, and can be seen as
  normalisation so that ? P(CF) 1
• hence P(F) can be omitted from classification
  calculations (since decision criterion is maximum
  likelihood)

28
Classification

• Bayesian classifier
• Extension to multiple features
• for simplicity, assume features are statistically
  independent, for each class (i.e. conditional
  independence)
• naïve Bayesian classifier
• compute class conditional probability, for each
  class, as product of class conditional
  probability for each feature
• Bayes rule P(CF1F2...Fn) ( P(F1F2...FnC) x
  P(C) ) / P(F1F2...Fn)
• ( P(F1C) x P(F2C) x ... x P(FnC) x P(C)
  ) / P(F1F2...Fn)

29
Classification

• Bayesian classifier numerical example
• Image segmentation
• image grey-level image of scene showing road,
  pavement and grass (e.g. for autonomous unmanned
  vehicle)
• known proportions of pixels in common scenes
  PavementGrassRoad 325
• problem classify each pixel in image captured by
  vehicle camera as road, pavement or grass
  using Bayesian maximum likelihood classification
• solution
• classifier Bayesian classifier
• feature F is pixel grey level
• for colour image, each colour component can be a
  feature (see naïve Bayesian classifier)

30
Classification

• Bayesian classifier numerical example (ctd.)
• Classifier training
• collect many typical images
• training set that contains representative data
• for each known class, select image areas
  containing pixels from class
• compute grey-level histogram for class
• normalise histogram so that sum of grey-level
  frequencies is 1
• i.e. estimation of class conditional
  probabilities (P(FC))
• estimate a priori probability for class (P(C))

31
Classification

• Bayesian classifier numerical example (ctd.)

Grey level
Label
Total
0
1
2
3
Pavement
15
12
26
28
81
Grass
30
20
6
4
60
Road
4
0
20
16
40
Training Grey-level histogram matrix
32
Classification

• Bayesian classifier numerical example (ctd.)

Grey level
Label
0
1
2
3
Pavement
15 / 81 0.19
12 / 81 0.15
26 / 81 0.32
28 / 81 0.35
Grass
30 / 60 0.5
20 / 60 0.33
6 / 60 0.1
4 / 60 0.067
Road
4 / 40 0.1
0 / 40 0
20 / 40 0.5
16 / 40 0.4
Training Class conditional probability matrix
33
Classification

• Bayesian classifier numerical example (ctd.)
• Training
• PavementGrassRoad 325
• hence, a priori class probabilities are
• P(Pavement) 3 / (3 2 5) 0.3
• P(Grass) 2 / (3 2 5) 0.2
• P(Road) 5 / (3 2 5) 0.5

34
Classification

• Bayesian classifier numerical example

Grey level
Label
0
1
2
3
Pavement
0.19 x 0.3 0.057
0.15 x 0.3 0.045
0.32 x 0.3 0.096
0.35 x 0.3 0.105
Grass
0.5 x 0.2 0.1
0.33 x 0.2 0.066
0.1 x 0.2 0.02
0.067 x 0.2 0.013
Road
0.1 x 0.5 0.05
0 x 0.5 0
0.5 x 0.5 0.25
0.4 x 0.5 0.2
Classification Scaled a posteriori class
probability matrix
35
Classification

• Bayesian classifier numerical example

Grey level
0
1
2
3
Grass
Road
Label
Grass
Road
Classification Class label matrix
36
Classification

• Classifier training
• Data-driven capture of parameters representing
  statistical distribution or syntactic
  configuration of salient class characteristics
• supervised training
• class labels used during training
• unsupervised training
• class labels not used during training (e.g.
  clustering)

37
Classification

• Classifier testing
• Testing estimation of recognition accuracy
• often uses real data simulation may be used
  (Monte Carlo)
• Accuracy measure
• error rate (often expressed as percentage)
• e.g. correct recognition rate, insertion rate,
  false acceptance rate, false rejection rate, ...

38
Optical Character Recognition
39
Optical Character Recognition
Generic OCR system
40
Optical Character Recognition

• Feature extraction methods
• Spatial domain to frequency domain transform
• Hartley, Fourier, or other transform
• Statistics
• mean, variance projection histograms
  orientation histograms

41
Optical Character Recognition

• Feature extraction methods
• Miscellaneous
• geometric measures
• ratio of width and height of bounding box, ...
• description of skeletonised characters
• graph description comprising line segments (e.g.
  strokes of Chinese characters)
• number of L,T, or X junctions, ...

42
Optical Character Recognition

• Feature extraction methods

Projection histograms
43
Optical Character Recognition

• OCR examples artificial neural network (1)

44
Optical Character Recognition

• OCR examples artificial neural network (2)

45
Optical Character Recognition

• OCR examples
• (see AALs book)

46
Other Recognition Applications

• Sample
• recognition of faces or facial expressions
• recognition of body movement (gestures, gait)
• recognition of handwriting (text, signature)
• industrial inspection
• autonomous vehicles, traffic monitoring
• ...
• (Exercise identify architectural components for
  these applications, and discuss factors affecting
  performance)

47
Summary

• Classical pattern recognition model
• pre/post-processing
• feature extraction
• classification
• Feature extraction representation of
  discriminant pattern characteristics
• Classification
• wide variety of classifiers
• supervised or unsupervised classifier training

48
Summary

• Components of generic OCR system
• image capture, image pre-processing, feature
  extraction, classification, post- processing
• Wide variety of features for OCR, e.g.
• frequency-domain representation
• statistical or geometric measurements
• skeleton descriptors
• ...