Statistical Pattern Recognition
Humans can achieve pattern recognition “at a glance” with little apparent effort. Much of pattern recognition is structural, being achieved essentially by analyzing shape. In contrast, statistical pattern recognition treats sets of extracted features as abstract entities that can be used to classify objects on a statistical basis, often by mathematical similarity to sets of features for objects with known classes. This chapter explores the subject, presenting relevant theory where appropriate.
• the nearest neighbor algorithm, which is probably the most intuitive statistical pattern recognition technique.
• Bayes’ theory, which forms the ideal minimum error classification system.
• the relation linking the nearest neighbor method to Bayes’ theory.
• the reason why the optimum number of features will always be finite.
• the ROC (receiver-operator characteristic) curve, which allows an optimum balance between false positives and false negatives to be achieved.
• the distinction between supervised and unsupervised learning.
• the method of principal components analysis and its value.
Statistical pattern recognition is a core methodology in the design of practical vision systems. As such, it has to be used in conjunction with structural pattern recognition methods and many other relevant subjects—as indicated in the heading of Part 4 of this volume. This is the context in which the face recognition ideas of Section 24.12 are included within the present chapter.