CS3361 Lecture 1

Pattern Recognition: 3/31/98

Bonus points are awarded for students who suggest useful links for these pages.

Readings:

• Chapter 24 (background on vision, interesting stuff on speech).
• Chapter 18 and 19: useful stuff on different types of pattern recognizers and how to train them.

Pattern Recognition

• printed character recognition [Assignment 1!]
• handwriting recognition [good project]
• speech recognition [good project]
• gesture recognition
• drawing recognition
• face recognition

Raw sensor signals are usually converted into features before recognition is performed.

Face recognition: sensor measures average color in pixels.

features: hair color, eye color, nose width, eye spacing, ...

Features are used in "Identikit" systems to generate drawings of faces.

Character recognition: binary sensor decides if pixel is black or white.

(binary scan 100dpi)

features: lines, blobs, holes, legs ???

Character recognition: grey level sensor averages light intensity in pixel.

(grey level scan 100dpi)

features: lines, blobs, holes, legs ???

Speech recognition: stereo microphones record sound waves versus time.

features: in 10 msec window: overall energy, energy in different frequency ranges, changes of features with respect to previous frame.

Handwriting and drawing recognition: pen strokes as a sequence of points (x, y, t).

pen stroke data corresponding to this picture

features: ???

Recognition Techniques

• Bayes' rule.
• Template matching.
• Simple parametric classifier.
• Decision trees.
• Neural networks.

Bayes' Rule.

derived from

P( class & features )
= P( class | features ) P( features )
= P( features | class ) P( class )

P( class ) can be calculated by dividing the number of items in a class by the total number of items.

P( features ) is a normalization and can be ignored if one is simply trying to maximize P( class | features ).

P( features | class ) is the probability that the observed features will be measured, given the class is known.