SSGD: Superpixels using the shortest gradient distance

As a pre-processing step for many problems in the field of computer vision, superpixel algorithms aim to over-segment the image by grouping homogenous pixels. In this paper, we propose a novel superpixel segmentation algorithm, namely Superpixel using the Shortest Gradient Distance (SSGD for short) in a κ-means clustering framework. Starting from initializing the superpixel seeds, bilateral filtering is applied to the texture-rich regions centered at initial seeds. Then, a novel distance function taking the shortest gradient distance into account is computed to enforce adherence to boundaries. Unlike using the simple Euclidean distance, the proposed combined distance function increases the accuracy of associating a pixel to a cluster. The experimental results demonstrate that our algorithm outperforms the state-of-the-art methods in this field. Source codes of SSGD are publicly available at http://sse.tongji.edu.cn/linzhang/ssgd/index.htm.