Saliency Detection Using Min-cut Proposals Framework
Meiling Sun, Fengxia Li, Sanyuan Zhao, Da Huo, Chenguang Yang · 2016
In saliency detection, almost all the approaches map an image into a graph and assign the saliency value to each element, e.g.pixel, region or superpixel.In this paper, we first utilize a series of image features among superpixels in the support vector machine (SVM) to train linear predicted models.For a well-performance model we take cross validation in the supervised learning.Then, we take the SVM regression models to predict initial saliency maps, while using SVM classifier to get the foreground and background seeds.Besides, we employ an objectness min-cut algorithm to obtain the segments of different proposals.Finally, after ranking these proposals, we select the top one integrating with the initial maps to achieve the final saliency maps.The proposed approach is tested extensively on four different databases and then compared with existing algorithms.