Covariance-based approach to texture processing

We present a simple and effective approach for texture processing that uses the eigenfeatures of local covariance measures. The covariance measures act as a texton encoder, producing texture code that is invariant to local and global textural rotations. This method uses only six features obtained from two scales of the invariant encoder to generate numerical representations for roughness, anisotropy, and other higher-order textural features. Classification results for synthetic and natural textures are presented. We discuss the effect of window sizes used at local and global scales on the performance of the classifier. (C) 1996 Optical Society of America