Research On Co-saliency Object Detection Algorithm Based On Hypergraph

There are two aspects in the field of salient object detection: one is salient object detection technology for single images and the other is co-saliency object detection technology for multiple images.Both of these technologies use a model or algorithm to make a computer identify and extract regions automatically in a natural image and form saliency or co-saliency maps.These regions conform to human visual characteristics and attract human visual attention.Salient object detection focuses on the salient objects of a single image.Co-saliency object detection is to extract areas that co-exist in a set of images and can attract human attention.These two technologies are important research in the file of computer vision.In recent years,because the salient object detection and cosaliency object detection has become widely used in image segmentation,image retrieval,co-segmentation,foreground annotation,and image redirection.Therefore,more and more scholars pay attention to them and propose a series of heuristic salient object detection algorithms and co-saliency object detection algorithms.However,existing salient object detection algorithms and co-saliency object detection algorithms have some problems: when the background of the detected image is complex,the foreground and background contrast are not sharp,or there are multiple salient objects in the images,cause the salient objects are incomplete,the boundaries of the salient objects are not clear and detect background as salient objects by mistake.In order to solve these problems,this paper puts forward the algorithm of salient object detection and co-saliency object detection based on hypergraph models and random walk and develops a co-saliency object detection system based on hypergraph models and random walk.(1)In the salient object detection algorithm based on hypergraph models and random walk,the simple linear iterative clustering algorithm is used to segment the superpixels firstly.Secondly,the fuzzy C-means clustering algorithm is used to build a general hypergraph model.Then,based on the prior knowledge of salient object detection,a weighted hypergraph model is constructed by using the location relationship and color similarity between the boundary superpixels and the center superpixels.This weighted hypergraph model is obtained by giving weight values to the vertices and hyperedges in the general hypergraph model.Besides,the transition probability matrix is obtained according to the rules of transition probability matrix generation.Finally,the transition probability matrix and the random walk algorithm are used to detect salient objects in a single image.(2)In the co-saliency object detection algorithm based on hypergraph models and random walk,this paper pre-processes each image in the group of co-images firstly.The pre-processing includes: using a saliency object detection algorithm based on hypergraph models and random walk to obtain the saliency maps of a single image,extracting the color features and global spatial features of each image,and merging the color features of each image.Secondly,using the fuzzy C-means clustering algorithm to cluster the merged color features to build a general hypergraph model.And a weighted hypergraph model is established by assigning weight values to the vertices and hyperedges in the general hypergraph model.The weight values are obtained by combining the three features of color similarity between regions in the co-images,the global spatial relationships of the single images,and the saliency values of the single images.Finally,the transition probability matrix is generated by using the transition probability matrix generation rule,and the transition probability matrix and random walk are combined to detect co-salient objects in co-images.(3)Design and develop a co-saliency object detection system based on hypergraph models and random walk.The salient object detection algorithm and co-saliency object detection algorithm based on hypergraph models and random walk and the step-by-step operation of these two algorithms is integrated into this system.The system has the functions of image superpixel segmentation,construction of weighted hypergraph models,and so on.The experiments show that the two algorithms proposed in this paper can extract the salient objects and co-salient objects in the images more completely and achieve satisfactory detection results.