关键词: 全变分流, 多变量混合泛化高斯分布（M2GGD）, 自然图像, 图像分割

Abstract:

An approach of natural image segmentation is proposed based on M2GGD distribution and total variation flow boundary. As the natural image is usually corrupted by some random noise, the finally segmented result with visual effect is poor. However, as edge information can distinguish out the non-homogeneous regions, therefore, an approach is designed by using total variation for extracting boundary, and meanwhile it is integrated with M2GGD distribution to improve the spatially constrained ability for natural image segmentation. For that the optimization problem of the designed energy function is NP hard, and then the region term and edge term is designed as t-link and n-link of multilayer graph cuts respectively, by employing maximum flow/minimum cut for optimization during the process of maximum expectation and maximum likelihood, and then, an approximate optimization solution is required. Finally, some noise corrupted synthetic image and natural scene image are adopted for experiment comparison and analysis，which demonstrates that the proposed approach with some superior advantages, such as anti-noise robustly, high quantified accuracy score, and the finally segmented results are more close to ground truth.

Key words: total variation, Multivariable Mixture Generalization Gaussian Distribution（M2GGD）, natural image, image segmentation