Towards Generalization of Image Anisotropy Analysis

Abstract

The texture anisotropy is a very important cue for object recognition. Suitable methods and procedures for the analysis of texture anisotropy in 2D images have been developed as early as in mid-eighties of the previous century. These methods were further improved during the early 2000th and successfully adapted for 3D (mostly medical) image analysis. This paper describes work in progress that refers to the third generation of the anisotropy analysis tool. Specifically, it presents a first attempt of generalizing the anisotropy estimation methods based on the very recent ideas of so-called classification gradient and related techniques. These techniques are capitalizing on an extended interpretation of the notion of image gradient treated as any quantitative difference between the two local pixel subsets measured by a suitable method. The approach is illustrated on detecting invisible anisotropy of 3D images of random textures caused by variation of high order statistics.