Improved Image Segmentation With A Modified Bayesian Classifier

A method for improving texture segmentation results by slightly modifying the decision surfaces of a Bayesian classifier is presented. Although a Bayesian classifier provides optimum classification within homogeneous regions, it does not necessarily provide accurate localization of region boundaries. In the proposed method, a modified classifier is formed by using a mixture probability density. This approach has the advantage that it is easily implemented in multidimensional classifiers such as those used in classifying the vector output of a filter bank. Experimental results demonstrate improved texture segmentation using the proposed classifier