A new method of texture segmentation

A new method to do texture segmentation is presented. We adopt the overcomplete wavelet packet frame to decompose the texture image into multichannel subimages. In all of the second-level subimages, only the one has the maximum variance among its same subchannel is selected as the feature image. Different feature extraction methods are applied to these 8 subimages according to the different character of themselves. Energy and entropy feature are extracted from the six detail subimages. In order to smooth the fragment caused by the detail feature, we extract the mean variance feature from the 2 approximate subimages. To improve the initial features, we proposed a quadrant mean filter to smooth the noise without over-blurring the boundary. The fuzzy c-means is provided to do classification of these features. The performance of this new method is demonstrated on the segmentation of Brodatz textures.