关键词: 显著性检测, 流形排序, 超像素分割, 多尺度分析, 图模型

Abstract: Visual saliency detection has many applications in object segmentation, adaptive compression, object recog-nition and so on. It has a challenge to accurately detect the most important regions from the nature images. This paper proposes a hierarchical saliency detection algorithm based on manifold ranking for the problem of low detection accuracy with the ignorance of the spatial layout information in the existing graph-based manifold ranking algorithms. Firstly, the super pixels by multi-scale analysis are done for the decomposition of the input image. Secondly, the boundary prior is used to compute the relevance ranking score between nodes by manifold ranking. Finally, by analyzing saliency cues from the multiple level graph model, the final saliency map is generated by combining the      saliency maps. The experimental results on ASD, CSSD, ECSSD and SOD image datasets, demonstrate that the detection precision outperforms the nine state-of-the-art algorithms while still preserving high recall.

Key words: saliency detection, manifold ranking, super-pixel segmentation, multi-scale analysis, graph model