A region-constrained asymmetric geodesic voting method for image segmentation

Image segmentation is a fundamental task in computer vision, aiming to accurately extract objects of interest from images in complex scenes. This paper presents a novel segmentation method based on a region-constrained voting scheme, integrating polar contour representation and an asymmetric geodesic voting framework. Classical geodesic voting methods rely on geodesic paths to approximate edge features but often fail to handle complex shapes or ambiguous boundaries robustly. Our primary contribution is the confinement of the voting process within a tubular neighborhood constructed around an initial contour. This spatial constraint effectively mitigates interference from irrelevant image regions and strengthens segmentation stability, particularly for objects with indistinct edges. We design a region voting score map with multiple source points, enabling better capture of complex and irregular boundaries. Furthermore, the PolarMask model is utilized to automatically generate multiple boundary-distributed key points, which are used as source points for geodesic path calculation, thereby significantly reducing the need for manual input and improving the method's applicability. By combining deep feature representation with the globally optimal geodesic tracing under an asymmetric quadratic metric, the proposed method achieves enhanced segmentation accuracy and robustness. Comprehensive experiments on medical and natural image datasets demonstrate that the proposed method substantially outperforms existing approaches, confirming its effectiveness in handling challenging segmentation tasks.