摘要：

已有基于图的流形排序显著性检测方法由于对背景先验假设过于理想,存在前景种子区域选取不准确的问题,从而影响检测效果。为此,提出一种新的显著性检测算法。通过计算图像的颜色增强Harris特征点,形成一个能够包含前景区域的凸包,并对图像进行超像素分割,计 算凸包内与凸包外所有超像素的差异度来选取前景种子。使用不同的区域特征描述子构建2个闭环无向图模型,分别将选取到的前景种子作为查询节点,采用经典的流形排序算法计算获得对应的显著图,融合2个显著图并作优化处理得到最终的显著图。在公开的MSRA10K和 ECSSD数据集上进行实验,结果表明,该算法在准确率、召回率、F值和MAE值等评价指标方面均优于传统的流形排序检测方法,其检测效果既能准确突出显著性目标,又能有效地抑制背景区域的干扰。

关键词: 显著性检测, 流形排序, 凸包先验, 显著图, 优化功能

Abstract:

Existing saliency detection method based on graph-based manifold ranking has the problem that the foreground seeds cannot be accurately selected because of idealistic prior background assumption,which influences the detection result.For this problem,a new saliency detection algorithm is proposed.The convex hull is got by calculating Harris feature corner of an original image to roughly estimate the location of salient region.Then,the original image is segmented into superpixels and the foreground seeds can be selected by calculating the dissimilarity between the superpixels inside and outside the convex hull.Two close-loop undirected graphs are constructed by using different regional feature descriptors.Accordingly,taking the selected foreground seeds as query nodes,two saliency maps are generated by adopting classical graph-based manifold ranking algorithm based on corresponding graphs.These two saliency maps are integrated and the result is further optimized to obtain a final saliency map.Experiments on the public MSRA10K and ECSSD data sets are done and the results show that the proposed method can outperforms the traditional graph-based manifold ranking algorithm on better precision rate,recall rate,F-measure and MAE value,and the detection result of the proposed method not only can accurately highlight the salient object in an image,but can suppresses the interference of the background area effectively.

Key words: saliency detection, manifold ranking, convex hull prior, saliency map, optimization function

中图分类号: 

• TP391