Statistical Pattern Recognition Part I: Clustering

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

• Introduction
• Features
• Clustering
• k-means clustering algorithm
• Pattern Similarity
• Summary

3
Introduction

• Examples of pattern recognition tasks

Image scene

• scene object (car, road, ...)

• alphabet letter (a, b, c, ...)
• writers name

Handwriting

• word (one, two, three, ...)
• speakers name

Speech
4
Introduction

• Pattern recognition
• Categorisation of patterns into a finite number
  of classes
• Common approaches
• statistical pattern recognition
• connectionist pattern recognition is sometimes
  considered as subset of statistical recognition
• syntactic pattern recognition

5
Introduction

• Pattern recognition
• Statistical
• classification based on statistical distribution
  of patterns
• Syntactic
• classification based on structural relationship
  between elements of pattern
• Connectionist
• classification using artificial neural networks
  (statistical basis?)

6
Introduction

• Terminology
• Feature(s) representation or description of
  salient pattern attribute(s)
• Class category , group, type
• patterns with common characteristics/properties
• Classification assignment of class label to
  observed pattern

7
Features

• Patterns often considered as points in space
• selected pattern attributes or properties
  (features) represented by
• numerical measurements
• linguistic labels
• space often multidimensional, hence
• set of numerical property values called feature
  vector

8
Features

• Examples attributes for
• pixel-based image segmentation
• pixel colour, edge strength,
• object recognition
• object shape, colour, size,

9
Features

• Example 2D feature plot for two image types (5
  samples of each)

10
Clustering

• Clustering unsupervised learning that identifies
  groups of similar data samples
• partial supervision (labelling of some samples)
  is possible
• Cluster group of similar data samples, which are
  dissimilar from samples in other groups
• for pattern recognition sample is synonymous
  with pattern

11
Clustering

• Example applications
• Image processing image segmentation into
  homogeneous regions
• data samples image pixels
• sample details pixel colour, edge strength, edge
  direction,
• Marketing market segmentation into groups of
  customers with similar buying behaviour
• data samples customers
• sample details database of purchases,
  demographics,
• Insurance profile identification of low- or
  high-risk clients
• data samples clients
• sample details database of client claim history,
  demographics,

12
Clustering

• k-means clustering algorithm
• Iterative algorithm for finding k clusters
• number of clusters assumed known
• Each cluster represented by prototype
• cluster centroid (mean)

13
Clustering

• k-means clustering algorithm pseudo code
• kMeansClustering(dataSamples, k)
• initialise cluster centroids
• do
• // assign each sample to a cluster
• for each data sample
• find nearest cluster centroid
• // update each cluster mean
• for each cluster
• update centroid
• until (centroids change is insignificant)

14
Clustering

• k-means clustering algorithm
• Limitations
• requires prior knowledge of number of clusters
• may fail to find optimal grouping of samples
• sensitive to choice of
• initial cluster prototypes
• e.g. what if a prototype is far from any data
  sample?
• distance measure (similarity measure)

15
Clustering

• k-means clustering algorithm
• As result of limitations, experimentation is
  often required
• multiple trials with different
• k
• initial prototype values
• distance measures

16
Pattern similarity

• Similarity often tied to geometrical distance
  between points that represent samples
• Many similarity measures are available
• choice is application dependent

17
Pattern similarity

• Minkowski metric
• popular similarity measure
• family of distance measurements between two
  points
• v point (represents feature vector)
• vk coordinate of point (i.e. feature)
• m dimensionality of data space (i.e. number of
  features)

18
Pattern similarity

• Minkowski metric (ctd.)
• special cases
• p 1 gives city-block distance (a.k.a. Manhattan
  distance or taxi-cab distance)
• equivalent to Hamming distance if observation
  vectors are binary
• p 2 gives Euclidean distance

19
Pattern similarity

• Minkowski metric pseudo code
• sum 0
• for feature from 1 to m
• diff featVector1feature featVector2feature
  
• pDiff power( abs(diff), p )
• sum sum pDiff
• sum power( sum, 1 / p )

20
Pattern similarity

• Minkowski metric (ctd.)
• limitations
• sensitive to non-uniform scale across features
• large-scale features may dominate distance
  calculation
• e.g. when features have different measurement
  units
• sensitive to correlation between features
• see limitations of minimum-distance classifier
  (in Part II)

21
Summary

• Clustering unsupervised identification groups of
  similar data samples
• Features salient pattern attributes
• k-means clustering algorithm
• iterative algorithm for finding k clusters
• each represented by prototype
• experimentation is often required as result of
  algorithm limitations
• Geometrical distance often used as similarity
  measure
• e.g. Minkowski metric