This chapter addresses the problem of segmenting textured images. Textured regions do not typically adhere to the piecewise-smooth or piecewise-constant assumptions that characterize most intensity-based segmentation problems. Julesz [82] pioneered the statistical analysis of textures and characterized textures as possessing regularity in the higher-order intensity statistics. This establishes the description of a textured image, or a Julesz ensemble, as one derived from stationary MRFs [189]. This principle forms the foundation of the proposed approach.
Image segmentation is one of the most extensively studied problems in computer vision. The literature gives numerous approaches based on a variety of partitioning criteria including intensity, color, texture, depth, and motion. The state-of-the-art in texture segmentation incorporates several important pieces of technology. One important component is the mechanism used to model or quantify the regularity in image textures. Researchers have developed progressively-richer descriptions of local image geometry [18,148,149] and sophisticated statistically-based metrics [39,87,124,83,144]; thereby capturing more complex distinctions between textures. Another area of focus, like in general image segmentation, concerns robust mechanisms for enforcing geometric smoothness on the segmented-region boundaries [118,153,117].
This chapter presents a method [4] that exploits the defining characteristics of a texture coupled with the generality of nonparametric statistical modeling. The method relies on an information-theoretic metric on Markov image statistics. The nonparametric modeling of the statistics of the stationary MRF imposes very few restrictions on the statistical structure of neighborhood intensities. This enables the method to easily adapt to a variety of textures. The method does not rely on a training stage and, hence, is unsupervised. These properties make it is easily applicable to a wide range of texture-segmentation problems. Moreover, the method incorporates relatively recent advances in level-set evolution strategies that use threshold dynamics [54,53].