Salient Object Detection Method Based On Visual Collaboration Perception

Salient object detection is one of the hot topics in computer vision.On the one hand,with the complication of the scene,it is difficult to accurately detect the salient object using single image information;on the other hand,the means of information acquisition is more and more diverse,which makes it possible to detect salient object using more comprehensive information.Therefore,in order to obtain more reliable results,this paper researches on salient object detection method based on visual collaboration perception,which using some properties of visual system such as multi-scale collaboration,three-dimensional perception etc.The research contents are as follows:1.In order to extract the co-saliency maps of image group with common salient object,we propose a low-rank multiscale fusion-based co-saliency detection method,which is inspired by visual multi-scale collaboration perception of scene information.The main ideas are as fellows: in the first stage,multi-scale segmentation of the image group is first performed,and the potential object regions are selected by the initial region salience value;then the potential object regions of each scale are clustered;and the weak co-saliency map of each scale is obtained through calculating the consistency of every region cluster.In the second stage,this paper introduces the low-rank constraint to calculate the weight of the weak co-saliency map for each image at each scale;Final co-saliency map is obtained by weighted fusion of multi-scale weak co-saliency maps.This method overcomes the limitation of the single-scale processing in the existing methods,and makes the co-salient object complete and accuracy.Its performance is verified on existing public datasets.2.To use the scene's comprehensive information to extract the salient object region of image,we propose a novel RGBD object detection method based on different-source image collaboration according to the characteristics of the visual three-dimensional perception system.The main ideas are as fellows: establishing different models of saliency map extraction for RGBD image and depth map respectively.More specifically,a short-connections based holistically-nested network is applied to extract RGB saliency map for RGB image;the depth saliency map is obtains through three kinds of saliency prior of depth image;finally,two kinds of saliency maps are multiplied to eliminate noise and heighten the common salient region;as result,we get the reliable RGBD saliency map.This method overcomes the difficulty of RGB image being easily affected by illumination and difficult to distinguish plane features in the traditional salient object detection.Comprehensive experiments over the public datasets demonstrate this method can effectively detect the salient object,and outperform other methods.The saliency map is more complete and the object boundaries are more accurate.3.In order to extract the co-saliency map of image group with comprehensive scene information,we propose a spatial-semantic RGBD co-salient object detection method based on the characteristics of visual multi-scale collaboration perception and visual three-dimensional perception of visual systems.The main ideas are as fellows: in the spatial collaborative processing channel,we use pixel-level clustering for RGBD image group,then spatial RGBD co-saliency map is obtained according to four kinds of co-saliency prior;in the semantic collaborative processing channel,these potential object regions of the RGBD image group are clustered,and semantic RGBD co-saliency map is obtained by the consistence of region depth range;finally,we obtain the final RGBD co-saliency map by adding two co-saliency maps up.This method extracts the co-saliency map in complex scenes through the collaboration of RGB image and depth map and the collaboration between images.This method effectively suppresses the cluttered background and overcomes the difficulty of small object detection.The experiment on two public datasets show that the proposed method has achieved advanced results.