| With the advent of the Internet age and the popularity of smart phones and other portable electronic devices, especially the rapid development of WeChat, twitter and social network, image is increasingly integrated into people’s life and affects our access to information greatly. Humans naturally and effortlessly judge the importance of image regions, and pay attention to important parts. How to make the computer to detect image saliency regions automatically via the simulation of human visual attention mechanism has become a significant research issue in the field of computer vision.This paper addresses saliency detection as a multi-view low-rank modeling problem. By extracting features in different modes from pre-generated superpixels, we achieve image features in multi-view which involve abundant information for saliency representation. We then develop a prior guided low-rank and sparse decomposition framework to obtain high quality saliency maps for given images. Different from conventional approaches which only use priors to generate feature matrix and solve certain existing low-rank models directly, we incorporate our newly defined priors into the saliency map learning process. Moreover, the (?)p,q norm regularization is introduced to a consistent saliency estimation based on features in multiple views. Extensive experiments and comparisons demonstrate that the proposed method can produce more precise and reliable results compared to state-of-the-art algorithms. |
