Combining Microscopic and Macroscopic Information for Rotation and Histogram Equalization Invariant Texture Classification

This paper presents a new, simple approach for rotation and histogram equalization invariant texture classification. The proposed approach is based on both microscopic and macroscopic information which can effectively capture fundamental intensity properties of image textures. The combined information is proven to be a very powerful texture feature. We extract the information at the microscopic level by using the frequency histogram of all pattern labels. At the macroscopic level, we extract the information by employing the circular Gabor filters at different center frequencies and computing the Tsallis entropy of the filter outputs. The proposed approach is robust in terms of histogram equalization since the feature is, by definition, invariant against flattening of pixel intensities. The good performance of this approach is proven by the promising experimental results obtained. We also evaluate our method based on six widely used image features. It is experimentally shown that our features exceed the performance obtained using other image features.