Coarse-to-fine Double Linear Iteration Method for Generating Superpixels

Abstract

To improve the performance of the superpixel segmentation algorithm and extract the edge information of the input image, a coarse-to-fine double linear iteration (CDLI) method for generating superpixels is proposed in this article. CDLI can be seen as an improvement of the SLIC (simple linear iterative cluster). In order to improve the accuracy of the superpixel segmentation algorithm, the density of the input image should be uneven. Regions with sparce edge information should have lower superpixel density while regions with rich edge information should have higher superpixel density. In order to achieve this goal, CDLI coarsely partitions the input image firstly, with a method which is similar to the SLIC. On the basis of rough segmentation, the region that needs further fine segmentation is selected. Finally, the selected region is partitioned with higher precision to form the final result. The entire process is equivalent to running SLIC twice. In a series of experiments, it shows that CDLI significantly improves the segmentation accuracy of the SLIC by nearly 1 percentage point while slightly reducing the time efficiency of the SLIC.