An improved algorithm for the segmentation of textured images is developed by means of feature selection and proposed grouping algorithm. First, the given image is devided into N by N blocks and the textural features for each block are estimated. Most important features for the given textured image are selected by Karhunen-Loeve expansion via pre-clustering and less important and noisy features which impede the separability of clusters are rejected. The proposed grouping algorithm is used for partitioning samples into a given umber of clusters on the smoothed histogram of feature vectors and thus the image is coarsely segmented. The fine segmentation of pixel unit using eight masks for homogeneity in a k by k window for only boundary blocks is presented. With the suggested algorithm, processing time (when except training phase) is reduced and accurate boundaries between texture regions are obtained in comparison with conventional algorithms. The simulation results on artificial and natural textured images show the agreement of this facts.
