A Clustering Approach for Color Texture Segmentation

The segmentation of natural images into perceptually uniform regions remains a challenging task. There is a substantial literature on color image segmentation, but there has been relatively little work on texture. A recently proposed perceptually-based color-texture segmentation algorithm achieves good performance on a wide variety of natural images. However, the computational requirements are quite substantial. The challenge, and primary goal of this thesis, is to develop computationally efficient algorithms without significant sacrifices in performance. The proposed approach is based on the observation that most natural textures consist of one or two dominant colors, and that the probability of having a third or even fourth dominant color is very low and can be negligible. We present a novel feature-aligned dominant-color-based clustering algorithm for image segmentation, which is aimed at segmenting natural scenes into perceptually uniform regions. The proposed approach uses the well-established adaptive clustering algorithm (ACA) to obtain spatial adaptive dominant colors, and relies on the two-dominant color representation of perceptually uniform textures to obtain a compact feature representation for each pixel in the image. At the core of the proposed algorithm, is a feature-aligned clustering algorithm, which is a generalization of ACA to include perceptually uniform textured regions in addition to regions of slowly varying colors. Like ACA, it includes adaptation to local image characteristics and spatial constraints in the form of Markov random fields. The initial step of the proposed feature-aligned adaptive clustering algorithm is a new feature-aligned K-means algorithm.;Experimental results with natural images indicate that the performance of the proposed approach is comparable to or better than the perceptually-based color-texture segmentation algorithm mentioned above, and comes at a drastically lower computational cost. We also show that the proposed approach outperforms other color-texture segmentation techniques.