Locating texture boundaries using a fast unsupervised approach based on clustering algorithms fusion and level set

Image segmentation deals with partitioning an input image into disjoint/non-overlapping regions. Among different segmentation algorithms, level set methods have been very popular. Less sensitivity to initialization, ability to split and merge the contour, and also, involving statistical inference have made level set even more accepted than similar methods like snakes. However, it is very time-consuming. To solve this problem, in this paper a fast variational approach is presented for texture segmentation. For this purpose, first a feature space based on non-linear diffusion is set up from CIE L*a*b* colour components. Then, this feature space is clustered by fusion of clustering algorithms. Finally, the produced cluster map is used in level set for contour evolution. As it is shown in the simulation results, our algorithm is robust in segmenting noisy texture. Also, it is faster than previous level set approaches for texture segmentation.