SUPERPIXEL SEGMENTATION BASED ON GRADIENT AND MANIFOLD

In today’s image processing tasks, the superpixel is often used as a method of dimensionality reduction for image as well as the basis of edge optimization. A super-pixel segmentation method based on gradient and manifold distance is proposed to solve the problem of experience-dependent segment number and discrete point of existing methods. It estimated the suitable number of superpixels for images adaptively, making segmentation for details more accurate and reducing over-segmentation for background. Experiments were conducted on BSDS500 dataset. We achieved good performance in various indicators. Especially, the elimination of discrete points leads to the compact with huge improvement.