Experiments in character recognition using linear and quadratic filters

This thesis describes the simulation of a character recognition system using rive filter designs based on probabilistic models of character patterns. Four of the designs yield linear filters. Of these, three are based on variations of a Gaussian model. The fourth is based on the assumption of independent binary-valued features. The latter design is shown to produce higher recognition rates than any of the others when tested on Munson's multi-author hand-printed characters. This filter design is also tested on two subsets of the Cornell machine-printed data base. The fifth filter design is a special case of a quadratic filter, based on a Gaussian model in which spatially stationary covariance statistics are assumed. This assumption results in a filter structure consisting of a linear operation on the pattern vector plus a linear operation on the autocorrelation vector of the pattern. This filter design is found to achieve lower performance than the best linear filter design when tested on Munson's characters, and nearly equal performance on the Cornell characters. However, there are indications that a filter of this structure could achieve higher performance for some choice of filter coefficients.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat