Abstract:

As multi-texture images are difficult to segment accurately,this paper proposes a texture segmentation method based on the complementary feature extraction of wavelet packet frame sub-band. In this method,first,the original texture image is decomposed by using the wavelet packet frame. Then,two sets of features in the neighborhood window of each pixel,namely,the average absolute deviation of sub-band coefficients and the mean and
standard deviation of the histogram of oriented gradients for sub-band coefficients,are extracted for each sub-band coefficient. Moreover,texture pixels are clustered via the improved spatial fuzzy c-means clustering. As the proposed method takes into consideration the spatial distribution of the local standard deviation of pixel feature values,it helps to obtain texture segmentation results with low misclassification rate of pixels near the texture boundary. Finally,several texture images from Brodatz album are used to test the proposed method. Fisher linear discriminant analysis indicates that the combination of the two sets of features is more effective in texture distinction than any single feature. Texture segmentation experiments show that the proposed segmentation scheme helps to achieve higher segmentation accuracy. In addition,test results of algorithm speed prove that the proposed method is applicable.

Key words: texture segmentation, wavelet packet frame, histogram of oriented gradients, spatial fuzzy c-means clustering