A novel segmentation approach for color images with progressive superpixel merging

We propose a novel approach to color image segmentation based on superpixels. We firstly apply normalized cuts algorithm to over-segment the images into small similar areas, namely, superpixels, which are local, coherent, and preserve most of the structure necessary for segmentation. Then the superpixels, as the elementary units, are progressively merged according to the merging cost, which consists of the dissimilarity, the continuity of boundary areas, and the compatibility between two adjacent superpixels. To enhance the color consistency under different illumination, the dissimilarity between two superpixels is measured by both the distribution probability and the context from R, G, B, and Saturation channels. The experiments demonstrate that our segmentation approach is effective and competitive, especially when the number of categories is small.