Multiscale asymmetry signatures for texture analysis

Abstract

This paper investigates model-based texture feature extraction from image multiscale representations. This approach offers a better texture characterization compared to using the classical energy output of a multiscale filterbank. The existing models assume symmetric density functions and we observe that this assumption is violated for some texture classes. This property is exploited to obtain improved texture characterization which can be used for texture classification, segmentation and retrieval.