Color image segmentation using iterative edge cutting, NUV-EM, and Gaussian message passing.

A new approach to image segmentation (grayscale or color) is proposed. It uses a (improper) Markov random field prior with sparsifying NUV terms (normal with unknown variance), which favors piecewise smooth images with sharp edges. The proposed algorithm iterates two steps. In the first step, the unknown scalar variances are learned by approximate EM (expectation maximization). The actual computations for this step boil down to iterative scalar Gaussian message passing, which scales linearly with the number of pixels. In the second step, all edges that were detected in the first step are cut and removed from further processing. Simulation results demonstrate that the proposed approach compares favorably with four standard methods.