Texture feature performance for image segmentation

This paper describes a comparative study of texture features, with particular emphasis on the applicability to unsupervised image segmentation. A benchmark test is introduced in which a set of 20 simple bipartite images, combining different stochastic textures separated by a stochastic boundary, is used for feature extraction and segmentation. The accuracy of the segmentation result, expressed in the mean boundary error, is used as an evaluation criterion. From the seven feature extraction methods tested, the Haralick, Laws and Unser methods gave best overall results. Results obtained also show that direct feature statistics such as the Bhattacharyya distance are not appropriate evaluation criteria if texture features are used for image segmentation. A small experiment on visual boundary tracking revealed that boundary error obtained here are similar to those obtained by machine segmentation. © 1990