A cellular automata approach for superpixel segmentation

This paper presents a novel Cellular Automata (CA) approach for image segmentation. We treat the image segmentation problem as cell merging in a cellular space constructed in the image plane. A cell is defined to be a pixel or a group of pixels with close RGB values. In each iteration, a cell checks the similarities between itself and its neighboring cells. Cells with similar properties are merged into large cells, which will eventually lead to high quality superpixels. The segmentation process is a trade-off between accuracy and computation cost. We have proved that the proposed approach is able to obtain satisfactory results efficiently while keeping image details.