Subsampling-based acceleration of simple linear iterative clustering for superpixel segmentation

Simple linear iterative clustering (SLIC) that partitions an image into multiple homogeneous regions, superpixels, has been widely used as a preprocessing step in various image processing and computer vision applications due to its outstanding performance in terms of speed and accuracy. However, determining a segment that each pixel belongs to still requires tedious, iterative computation, which hinders real-time execution of SLIC. In this paper, we propose an accelerated SLIC superpixel segmentation algorithm where the number of candidate segments for each pixel is reduced effectively by exploiting high spatial redundancy within natural images. Because all candidate segments should be inspected in order to choose the best one, candidate reduction significantly improves computational efficiency. Various characteristics of the proposed acceleration algorithm are investigated. The experimental results confirmed that the proposed superpixel segmentation algorithm runs up to about five times as fast as SLIC while producing almost the same superpixel segmentation performance, sometimes better than SLIC, with respect to under-segmentation error and boundary recall.