||In this thesis, we propose a new color segmentation approach that combines the advantages given by the reliability of global color statistics and the flexi-bility of local image compositing, and preserves the necessary spatial and color coherence in the image. Global and local information are integrated by belief propagation on a Markov Random Fields (MRF) network to identify relevant neighborhood information for optimizing the probabilistic grouping of image pix-els. The global color statistics of an image is represented by a Gaussian Mixture Model (GMM), whereas color of a local pixel is explained by a local image com-positing model. Each pixel color consists of a mixture of constituting colors weighted by the membership probabilities corresponding to the Gaussians of the converged GMM. Transparency is thus naturally introduced in this probabilistic framework, and we call our approach soft color segmentation due to the presence of fractional or soft region boundaries. To adequantly consider pertinent global and local information within the same framework, an alternating optimization scheme is proposed to iteratively solve for the global and local parameters. Our iterative method is fully automatic and empirically shown to converge to an optimal solution subject to the necessary spatial and color coherence. We perform quantitative and qualitative evaluation on our method, and compare our results with those generated by state-of-the-art techniques in automatic color segmentation and user-assisted image matting. Our segmentation results are comparable or more preferred in all the cases we tested, leading to a wide range of potential applications under the uniform framework: natural image matting, color transfer, image deblurring, and colorizing greyscale images.