The Co-saliency Detection Of RGBD Images Based On Quaternion And Hypergraph

The saliency detection and the co-saliency detection are two aspects in the field of saliency detection.The former is to operate on a single image,the latter is to operate on a set of common image groups.At present,the saliency and co-saliency detection for RGB images is becoming more and more mature,and has been widely used in image segmentation,image retrieval and co-segmentation.When the background of the detected image is complex and the contrast between the foreground and the background is not high,the existing results have the problems of incomplete detection of salient objects and unclear boundaries.With the development of imaging technology,the depth information in RGBD images has been proved to play an important role in object segmentation and saliency detection and other computer vision tasks.Hence,this paper is to perform cosaliency detection on RGBD images and use the depth information in RGBD images to enhance the recognition of co-saliency.And it uses adaptive super pixel division technology,Hu moment feature and Hypergraph model to detect RGBD images.(1)To implement the adaptive super-pixel division technology for RGBD image.Firstly,the RGBD image is divided into non-overlapping blocks.Secondly,a new similarity matrix is constructed according to the color information and depth information in the RGBD image.Thirdly,the entropy in the RGBD image is calculated to adaptively determine the threshold.Finally,the availability function and the responsibility function are used for affinity propagation to segment the RGBD image adaptively.(2)To construct a new quaternion structure and extract the Hu feature of the RGBD image.Firstly,the depth information in RGBD image is used to replace the real part of quaternion,and the color information in RGBD image is used to replace the imaginary part of quaternion.Secondly,the constructed quaternion is calculated to get Hu moment features.(3)To implement the co-saliency detection of RGBD images based on quaternion and hypergraph.Firstly,the image is divided into the super-pixels according to the adaptive super-pixel division technology.Secondly,the features are extracted of each super-pixel in RGBD image.Thirdly,a new hypergraph model is constructed for RGBD images.Finally,random walk is performed on the newly constructed weighted hypergraph model to realize the co-saliency detection of RGBD images based on quaternion and hypergraph.Experiments are carried out on two RGBD co-saliency datasets and the proposed method based on quaternion and hypergraph works well in co-saliency scenes.It can extract saliency and co-saliency regions completely,and the boundaries of the extracted saliency and co-saliency regions are relatively clear.