Generic Image Segmentation in Fully Convolutional Networks by Superpixel Merging Map

Jin-Yu Huang, Jian-Jiun Ding; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

Recently, the Fully Convolutional Network (FCN) has been adopted in image segmentation. However, existing FCN-based segmentation algorithms were designed for semantic segmentation. Before learning-based algorithms were developed, many advanced generic segmentation algorithms are superpixel-based. However, due to the irregular shape and size of superpixels, it is hard to apply deep learning to superpixel-based image segmentation directly. In this paper, we combined the merits of the FCN and superpixels and proposed a highly accurate and extremely fast generic image segmentation algorithm. We treated image segmentation as multiple superpixel merging decision problems and determined whether the boundary between two adjacent superpixels should be kept. In other words, if the boundary of two adjacent superpixels should be deleted, then the two superpixels will be merged. The network applies the colors, the edge map, and the superpixel information to make decision about merging suprepixels. By solving all the superpixel-merging subproblems with just one forward pass, the FCN facilitates the speed of the whole segmentation process by a wide margin meanwhile gaining higher accuracy. Simulations show that the proposed algorithm has favorable runtime, meanwhile achieving highly accurate segmentation results. It outperforms state-of-the-art image segmentation methods, including feature-based and learning-based methods, in all metrics.