Image Saliency Detection Based On Multi-graph Model And Manifold Ranking

Image saliency detection is to use computer to simulate human visual attention mechanism and extract the interesting regions from the input image for subsequent analysis.Image saliency detection has a wide range of applications,including target recognition,image retrieval,image segmentation,content-aware image compression,and so on.Most of saliency detection algorithms can detect salient objects in images with a simple background,but images with a complex background,the detected salient objects still have such problems,for example,the internal of salient objects are not uniform,and objects are not complete.To deal with these problems mentioned above,we propose an image saliency detection algorithm based on multi-graph manifold ranking.The main work and novelty of this thesis are as follows:This thesis proposes an image saliency detection algorithm named image saliency detection based on multi-graph manifold ranking.First,it constructs K regular graph and KNN graph respectively with superpixels as nodes;then,based on the global foreground hypothesis,the saliency values of nodes in both of the graph models are calculated by the manifold ranking algorithm;final,the two graph models are fused to obtain the saliency map.The proposed algorithm uses two graphs to construct the relationship among data nodes,which makes up for the incompleteness of the salient object obtained by using single graph model and suppress background noises to a certain extent.In this thesis,the image saliency detection based on multi-graph manifold ranking is extended on multi-scales.Because there are a few detail information in low-scale image,which can emphasize important objects and suppress noise,the low-scale image is constructed from the input image and the saliency map is obtained by using multigraph manifold ranking on different scales.Then the final saliency map is obtained by fusing the multi-scale saliency maps.We compared the proposed algorithm with 14 state-of-the-art algorithms on three public databases including MSRA-10 K,SED2,and ECSSD.We use PR curve,ROC curve,F-measure,and AUC value to evaluate different algorithms.The experimental results show that the algorithms proposed in this thesis can extract the salient objects quickly and accurately,and the salient objects are more complete,and the internal of the salient objects is more uniform,especially in images which have complex backgrounds.