Gestalt-grouping based on path analysis for saliency detection

Abstract

Due to the arbitrary scales, uncertain distributions of objects and cluttered background in natural scenes, uniformly detecting salient regions remains a challenge. This paper first proposes a Gestalt-grouping connectedness method based on path analysis to reflect the topological relationship between image pixels. Inspired by the Gestalt principles of feature grouping, we apply a smoothest path-based distance metric to capture the similarity, local proximity and global continuity between image pixels. The distance is small if the image pixels belong to the same visual region and large otherwise. To identify salient regions in natural images, we then propose a path-based background saliency model that integrates both the topological connectedness and appearance dissimilarity. Experimental results demonstrate the advantage of applying the path-based background saliency model in uniformly highlighting salient regions in images with complex backgrounds.