Lecture Twentytwo Pattern Recognition

1
Lecture Twenty-twoPattern Recognition
2
Introduction

• The purpose of pattern recognition is to place
  objects in a given world into categories
• The interface between the world and the pattern
  recognition system is provided by sensors

3
Pattern Recognition Procedure

• The first step of the procedure extracts features
  from the input data which characterise the
  objects.
• Based on these features, the objects are
  identified and sorted into classes.

4
Labels

• In order to sort the objects, the system needs
  information concerning the features of the
  objects i.e. the system needs a label

5
Types of Pattern Recognition System

• Un-Supervised
• These are able to generate their labels
  themselves, assigning them to objects with
  similar features which could belong to the same
  class. These systems cannot recognise the object
  they are analysing.

6
Types of Pattern Recognition System

• Supervised
• These systems are taught such information as
  this is a banana. The work for this system is
  divided into two stages
• Training step - requires a teacher who describes
  the properties of each class.
• Classification step - compares the features of an
  actual object with those values which have been
  taught. The object is assigned to the class
  which fits best.

7
Problem with second approach

• If someone inserts a foreign object into a fruit
  recognising system (e.g. a calculator), then
  there is a class into which the calculator fits
  better than any other. This can be overcome by
  introducing a rejection level which tests the
  limits of similarity.

8
Feature Space

• The features extracted from the input data form a
  feature space

colour
blue
Plum
green
Apple
Banana
yellow
Orange
red
compactness
(surface areaVolume)
9
Problems with the Feature Space

• Too few or unsuitable features results in classes
  which are not separable
• If an appropriate choice of features is not
  possible, or too expensive, then the aim should
  be to use features leading to a minimum
  classification error.
• To avoid the classification errors, it may be
  necessary to reduce the world, e.g. restrict
  the colour of apples to green. If this is not
  practical, then an additional feature must be
  introduced,e.g. surface texture.