利用层次先验估计的显著性目标检测

有效的显著性目标检测在计算机视觉领域一直是具有挑战性的问题.本文首先对图像进行树滤波处理,采用Quick shift方法将其分解为超像素,再通过仿射传播聚类把超像素聚集为代表性的类.与以往方法不同,本文提出根据各类中拥有的超像素的类内和类间的空间离散程度及其位于图像边界的数目,自适应地估计先验背景,并提取条状背景区域;由目标性度量(Objectness measure)粗略地描述前景范围后,通过与各类之间的空间交互信息,估计先验前景;再经过连通区域优化前景与背景信息.最后,综合考虑各超像素与先验背景和前景在CIELab颜色空间的距离,并进行显著性中心加权,得到显著图.在MSRA-1000和复杂的SOD数据库上的实验结果表明,本文算法能准确、完整地检测出显著性目标,优于21种State-of-the-art算法,包括基于部分类似原理的方法.

Exploiting Hierarchical Prior Estimation for Salient Object Detection

Effective salient object detection is still a challenging problem in computer vision. In this paper, images are processed by tree filter firstly. Then quick shift is adopted to decompose images into perceptually homogeneous superpixels. This is followed by using affinity propagation clustering to aggregate all the superpixels into representative clusters. Different from previous methods, this paper proposes a novel adaptive background prior estimation strategy. The intra-cluster and inter-cluster spatial variances of superpixels owned by some cluster are calculated, and the numbers of superpixels located along the image boundary are counted to complete the processing. Also, the strip regions of background are extracted. The objectness measure is employed to get a coarse foreground scope, which is then used to compute the spatial interactive information with all the clusters, so as to get the foreground prior. After the optimization based on connected regions, A final foreground and background prior are confirmed. A saliency map is generated by measuring the differences of CIELab color space between all the superpixels and the background and foreground prior, enhanced by salient center weighting. Experimental results on MSRA-1000 and complicated SOD databases show that the proposed method can accurately detect the whole salient object. It is superior to the 21 state-of-the-art methods, including the methods partially based on similar principles.