Three-dimensional structural texture modeling and segmentation

Abstract

The structural approach, one of the various modeling techniques in texture analysis, perceives texture as a region composed of detectable primitive elements/tokens. In this paper, we extend this concept to three-dimensional (3D) image and develop a segmentation scheme based on clustering of 3D texture tokens. 3D structural texture is composed of texture tokens which are modeled as tiny objects embedded inside image sequence. Texture tokens located spatially close to one another and having similar shape are grouped into homogenous texture volume. Some of the texture tokens may be in contact with one another to form clump-like structures. Our approach first disconnects these clumps by a 3D clump decomposition process. A 3D irregular pyramid construction process is then carried out on these isolated 3D tokens. The process inputs a Delaunay graph with each vertex representing the shape feature vector of a texture token and each edge representing the neighborhood relationship of two texture tokens. Through stepwise grouping of similar shape tokens, a segmentation of the texture volume is obtained.