一种基于图的流形排序的显著性目标检测改进方法

摘要

To overcome the shortage that the spatial connectivity of every node is modeled only via thek-regular graph and the idealistic prior background assumption is used in existing salient object detection method based on graph-based manifold ranking, an improved method is proposed to increase the precision while preserving the high recall. When constructing the graph model, the affinity propagation clustering is utilized to aggregate the superpixels (nodes) to different color clusters adaptively. Then, based on the traditionalk-regular graph, the nodes belonging to the same cluster and located in the same spatial connected region are connected with edges. According to the boundary connectivity, the superpixels along the image boundaries are assigned with different background weights. Then, the real background seeds are selected by graph cuts method. Finally, the classical manifold ranking method is employed to compute saliency. The experimental comparison results of 4 quantitative evaluation indicators between the proposed and 7 state-of-the-art methods on MSRA-1000 and complex SOD datasets demonstrate the effectiveness and superiority of the proposed improved method.%该文针对现有的基于图的流形排序的显著性目标检测方法中仅使用 k-正则图刻画各个节点的空间连接性的不足以及先验背景假设过于理想化的缺陷，提出一种改进的方法，旨在保持高查全率的同时，提高准确率。在构造图模型时，先采用仿射传播聚类将各超像素(节点)自适应地划分为不同的颜色类，在传统的k-正则图的基础上，将属于同一颜色类且空间上位于同一连通区域的各个节点也连接在一起；而在选取背景种子点时，根据边界连接性赋予位于图像边界的超像素不同的背景权重，采用图割方法筛选出真正的背景种子点；最后，采用经典的流形排序算法计算显著性。在常用的MSRA-1000和复杂的SOD数据库上同7种流行算法的4种量化评价指标的实验对比证明了所提改进算法的有效性和优越性。