层次图融合的显著性检测

Saliency Detection Based on Hierarchical Graph Integration

显著性目标检测在物体分割、自适应压缩和物体识别等领域有众多应用,从自然场景中准确检测出最重要的区域一直是个挑战。针对现有的基于图的流形排序算法,因忽略特征的空间信息而导致检测准确率不高的问题,提出了一种基于流形排序的多尺度显著性检测算法。首先对原始图像进行多尺度下的超像素分解。然后利用边界先验,根据流形排序算法计算查询点与其余结点的相关度排序。最后通过构建图模型,从多层结构中分析显著性线索,对显著图进行融合得到最终结果。在ASD、CSSD、ECSSD和SOD数据集上,同9种流形算法进行对比实验,结果表明该算法在保持高查全率的同时也提高了准确率。

Visual saliency detection has many applications in object segmentation, adaptive compression, object recognition 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.