Research Of Salient Object Detection Based On Prior Integration And Feature Guidance

As one of the most important part in computer vision,salient object detection attempts to detect the salient region in the image accurately.It has a wide applications,such as image segmentation,image compress,object detection and recognition,image indexing et al.The human visualization system could recognize the salient object in the image in a rapid and efficient fashion,this mainly owe much to different regions could provide different levels of stimulus.Whereas,salient object detection try to study this characteristic and recognize the salient regions in the image according to the quality difference perceived in different regions.Salient object detection mainly can be divided into two streams,i.e.top-down and bottom-up,respectively.Top-down methods utilize low-level clues such as the color feature of image,characteristic of spatial distance,texture feature to deal with the salient object detection problem.Due to the lack of guidance of upper knowledge,bottom-up methods always with the problem of false detection.Thus,we propose the prior fusion based global and local salient region detection algorithm.Some of the traditional ccontrast prior based methods only use the global contrast prior,this may lead to massive error detections of background region;while some only utilize the local contrast prior,it may fail to uniformly highlight the salient regions.Target to solve,above mentioned issues,we first propose to consider global and local views simultaneously and make those two methods complementary to each other.At the same time,this paper further improved the robustness and effectiveness of our algorithm based on the global and local views through prior fusion method.From the perspective of global views,we fuse the convex-hull based centerness prior in the global comparative prior,through the estimation of rough style of salient object regions to solve the error detections in the scarce background regions.From the local perspective,we introduce the compactness prior in the local contrast prior,and detect the salient object via the computation of spatial compactness of super-pixels in the images.Thus,we could solve the issue of the salient regions but cannot be uniformly highlighted.Aiming at suppressing the background regions and uniformly highlight salient regions,we also utilize the smoothness prior to optimize the salient regions from the perspective of the global and local angles.Although with the occurrence of robust distance measure sensors and depth features becoming more and more import in the area of navigation and robotics.The multi-modal RGB-D images are drawing more and more attention than only RGB image.However,existing models for salient object detection either works good on RGB images but cannot used directly on the domain of RGB-D images,or utilize the feature of RGB and depth feature independently and combine the results in a violent manner without adaptive fusion.In this paper,firstly,we considered the adaptive fusion of RGB and CIELab color features,and propose the feature fusion based manifold ranking algorithm for salient region detection of RGB image.Hence,we could guide the saliency detection of RGB image with the help of depth feature and further improve the detection results.Secondly,depth feature based preliminary saliency maps and depth feature information,we propose OD algorithm to improve the saliency results of RGB image.Thirdly,according to the relations between salient detection results of RGB images and depth feature information of depth images,we propose the S-D probability correction method for efficient correction of preliminary detection results of depth images.The correction operation will guide the salient object detection of depth images more efficient on the basis of RGB saliency detection results,thus it could efficiently improve the saliency detection performance of depth images;then we propose the OC algorithm based on the saliency results of RGB images and CIELab color feature information to further improve the salient object detection results in a large margin.We evaluate the saliency detection results on three benchmarks,i.e.MSRA-1000,CSSD and ECSSD.And extensive experiments demonstrate that the proposed algorithm improved the performance of salient object detection in a large margin compared with existing state-of-the-art approaches.Experiments on RGBD-1000 public dataset demonstrate that the proposed algorithm obtained huge improvement compared with existing state-of-the-art methods.