Contour Extraction by Mixture Density Description Obtained from Region Clustering

Abstract

This paper describes a contour extraction scheme which refines a roughly estimated initial contour to outline a precise object boundary. In our approach, mixture density descriptions, which are parametric descriptions of decomposed sub-regions, are obtained from region clustering. Using these descriptions, likelihoods that a pixel belongs to the object and its background are evaluated. Unlike other active contour extraction schemes, region-and edge-based estimation schemes are integrated into an energy minimization process using log-likelihood functions based on the mixture density descriptions. Owing to the integration, the active contour locates itself precisely to the object boundary for complex background images. Moreover, C^1 discontinuity of the contour is realized as changes of the object sub-regions' boundaries. The experiments show these advantages.