Image Saliency Detection Based On Graph Model Strategy

The human vision system can always naturally extract the most informative object or relevant area in the scene,even for various and complicated images.Image saliency detection,as an important branch of image processing,analyzes and collects image information by simulating the mechanism of the human vision system in order to extract the most salient areas.This technology is not only the basis for many high-level applications such as object recognition and image retrieval,but also is widely used in many fields such as communications and military.Algorithms based on graph models have become more popular in recent years because graph-based models are concise and efficient in expression of image features.In the existing algorithms,the graph-based often has limitations in the expression of the difference between the foreground and the background.The detection results lead to the lower gray value of the salient object region and the inaccurate edges of the salient object.Aiming at the above limitation,a novel image saliency detection algorithm based on graph model strategy is proposed in this thesis.The main work of the thesis is in the following four aspects:(1)According to the two principles of object compactness and regional homogeneity,saliency region extraction is performed on the image.Different from the traditional regular graph structure,the two-layer sparse graph structure is used to extract the compactness of the nodes,and redundant edge connections among many nodes are removed in the proposed algorithm.The information of the nodes in the object area is prominently revealed by means of node propagation and diffusion.In addition,a novel regional homogeneity graph structure is constructed with heterogeneous differences between potential foreground and potential background regions,where uncertain nodes are adjacently classified by location feature and color feature.Besides,with the salient information presented by the object compactness,different saliency confidences are given to the seed column vectors to improve the accuracy of detection.(2)Improvement of single-layer cellular automata based on the centrality of complex network is performed.In the proposed algorithm,a graph-based theoretical derivation of the diffusion matrix is performed so that the corresponding eigenvectors are visually encoded.The eigenvalues and eigenvectors with strong influence are selected and remained for the purpose of integration into a new type of propagation matrix to replace the original influence factor matrix.And the re-integrated diffusion matrix is used to calculate the initial value of the saliency diffusion in the cellular automata.In addition,the centrality of complex network is introduced to give the consistency matrix a strong saliency discrimination power,and then the improved single-layer cellular automata diffusion algorithm is used to obtain a saliency detection effect better than that of multi-layer diffusion.(3)According to the principle of consistency,the saliency results are fused with the graph-based model strategy.The saliency extraction stage based on object compactness and regional homogeneity expresses the difference between the foreground and background of the image from the nodes of the graph structure;The improvement stage of single-layer cellular automata based on the centrality of complex network separates foreground and background from the diffusion of strong influence eigenvalues and eigenvectors.Therefore,in order to integrate the advantages of multiple graph-based model for saliency detection,the color space of 11 color names is introduced,and the saliency maps of the above two stages are fused according to the principle of consistency in detection region.(4)Multi-stage experiments and results analysis of the algorithm are performed.According to the algorithm flow,the validity of multiple graph models is verified in turn.Compared with 13 algorithms on public datasets,the saliency detection results are evaluated qualitatively and quantitatively,which are used to prove the superiority of the proposed algorithm.