Feature analysis for scaled and rotated texture segmentation

In this paper a class of spatial and spectral features obtained from the most popular methods for texture classification are analyzed for their ability to segment scaled and rotated textures. The features are defined from: the co-occurrence matrix using spatial-gray level dependence method, the gray level run length method, the spatial gray level difference method, the Fourier spectrum and the Gabor transform. Several experimental results are presented and discussed. The Gabor features and the co-occurrence matrix features give best results followed by the pixel-based, Fourier and spatial gray level difference features. The run length features perform poorly. The feature analysis shows that the angle at which the second order statistical features such as the co-occurrence features and gray level difference features are obtained, control the performance of the features. The Gabor features gives best results when the filter is tuned well to the orientation of the textured images. The features can be used for segmenting images obtained at different resolutions and transformations by augmenting the statistical invariance of the features to scale and rotation. Contrary to earlier conclusions, the analysis of these feature sets for scaled and rotated texture segmentation proves that some of the features can very well be applied to practical situations and further enhanced for better performance.