Stochastic color image segmentation using spatial constraints

This paper describes an automated method for segmenting color images based on a modified stochastic region merging strategy with multi-scale spatial constraints. First, a bilateral decomposition is performed, and an over-segmentation process is then performed based multichannel information and multi-scale gradients. Next, each sub-region is represented using a normalized color histogram in the CIE L*a*b* color space, and a region adjacency graph is constructed based on the over-segmentation results. Finally, a stochastic region merging strategy with spatial constraints is performed on the region adjacency graph to construct one segmentation map for each scale of representation. Our preliminary visual and quantitative experimental results on the Berkeley image database (BSDS500) are encouraging, and suggest that our proposed approach can provide accurate segmentation results.