Improved Feature for Texture Segmentation Using Gabor Filters

The local structure of texture can be obtained by transforming a texture image to new basis given by convolving it with Gabor filters in order to segment images contain multiple textures. In recent years, some features have been proposed, but the segmentation performance can still be improved. In this paper, an improved energy feature, which using variable window size decided by scale of Gabor kernel, has been proposed. So the local properties in an appropriated neighbourhood can been captured better. Since we focus on observing the performance of new features, we use PCA (principal component analysis) as the dimension reduction method and K-means algorithm as clustering algorithm for simplicity. From the experimental results using several features, it can be seen that our feature can improve the separability of texture boundaries and irregular textures.