A fast superpixel algorithm with biased-clustering using visual saliency

作者：Li, Peng; Yang, Yang; Fang, Tao

来源：Journal of Xi'an Jiaotong University, 2015, 49(1): 112-117 and 138.

DOI：10.7652/xjtuxb201501019

• Biased-clustering
• Distance functions
• Global clustering
• Initialization step
• Non-uniform mesh
• Over segmentation
• Visual Attention
• Visual saliency

A novel biased-clustering superpixel algorithm is proposed in the framework of SLIC to improve the problem that the conventional superpixel methods have bottleneck in controlling the tradeoff between superpixel number and boundary adherence. The algorithm employs the visual saliency into the non-uniform mesh initialization step and the biased-clustering distance function by noticing that human's visual attentions are distinctive to different salient objects, so that dense over-segmentations are generated in salient regions to keep sufficient information for object's boundary;while sparse segmentations are generated in non-salient regions to reduce the number of segmentation blocks. Moreover, the ideas of one-step global clustering and gradual boundary refining are applied to speed up the algorithm. Experimental results and comparisons with several state-of-the-art superpixel algorithms show that the proposed algorithm reflects the boundary adherence for salient objects well under the same number of segmentation blocks, and has a higher boundary recall and lower under-segmentation, as well as the least time-consuming.

• 出版日期2015
• 单位上海交通大学