Understanding Natural Texture

Summary: Texture analysis is a tool to detect the forest from the trees. Understanding texture does not solve the image understanding problem, nor does it produce the ultimate region grower. Texture descriptions provide a useful input to higher-level understanding programs, and a valuable cue for many segmentation problems. We have presented here a technique for producing useful descriptions of texture based on local shape statistics of simple regions called texture primitives, and on a simple set of relationships among similar classes of texture primitives. The usefulness of this description has been demonstrated through standard discrimination/classification tests. More importantly, these descriptions provide quantification of discrimination criteria and allow recognition of similar textures over transformations of scale, intensity, and rotation. Some different textures can be normalized to produce descriptions which are as similar as possible, and these normalized textures do in fact appear quite similar. Proposed extensions include using a world model to detect spatial orientation and distance transformations