Topology Preserved Regular Superpixel

Abstract

Most existing super pixel algorithms ignore the topology and regularities, which results in undesirable sizes and location relationships for subsequent processing. In this paper, we introduce a new method to compute the regular super pixels while preserving the topology. Start from regular seeds, our method relocates them to the pixel with locally maximal edge magnitudes. Then, we find the local optimal path between each relocated seed and its four neighbors using Dijkstra algorithm. Thanks to the local constraints, our method obtains homogeneous super pixels with explicit adjacency in low-texture and uniform regions and, simultaneously, maintains the edge cues in the high contrast and salient contents. Quantitative and qualitative experimental results on Berkeley Segmentation Database Benchmark demonstrate that our proposed algorithm outperforms the existing regular super pixel methods.