Monte Carlo cluster refinement for noise robust image segmentation

This paper explores a stochastic approach to refining clustering results for data with spatial-feature context such as images under the presence of noise. We formulate the clustering problem as a maximum a posteriori (MAP) problem, and refine clustering results using importance-weighted Monte Carlo posterior estimates based on between-neighborhood error statistics to account for local spatial-feature context within a global framework. This cluster refinement approach is non-iterative and can be integrated with existing clustering methods to achieve improved clustering performance for image segmentation under high noise scenarios. Experiments on synthetic gray-level images, real-world natural images, and real-world satellite synthetic aperture radar imagery illustrate the proposed method’s potential for improving clustering performance of existing clustering algorithms for image segmentation under high noise situations.

► We explore a stochastic cluster refinement approach for image segmentation.
► We formulate the clustering problem as an maximum a posteriori (MAP) problem.
► We refine clustering using importance-weighted Monte Carlo posterior estimates.
► Experiments show improved clustering for image segmentation under high noise.