Salient Object Detection In RGBD Images Based On Hypergraph Model And CNN

Salient object detection aims to use computers to simulate the working mechanism of human vision,ignore the invalid regions in the visual scene,and extract the most informative regions in the image accurately and quickly.Nowadays,salient object detection is mainly divided into two types of methods based on feature extraction and deep learning.The former uses manual features of images for salient object detection,while the latter uses deep semantic features of images for salient object detection.Due to the rapid development of imaging technology,the depth information contained in RGBD images can effectively enhance the representation of salient objects in images,and thus has been widely used in computer vision tasks such as salient object detection.However,for some images with complex background and low foreground and background contrast,the existing salient object detection methods will have the problem of incomplete salient object detection and unclear boundary.From the perspective of feature extraction and deep learning,we propose two models of RGBD image salient object detection:Firstly,We propose a salient object detection method for RGBD images based on superpixel and hypergraph.Aiming at the problem that the existing superpixel segmentation methods can not fit the weak edge of the image,and even have the wrong segmentation,we first propose a novel superpixel segmentation method for RGBD images.Secondly,a weighted hypergraph model is constructed based on the color feature,optimized depth feature and global spatial feature similarity of image superpixels.Finally,the random walk algorithm is used to rank the importance of superpixels to get salient object detection results.The experimental results show that the hypergraph model salient object detection method constructed in this thesis is superior to the feature extractionbased method in the comparison method,but compared with the existing deep learning method,there are still shortcomings in the detection accuracy.Secondly,a salient object detection method based on gated cross-modal attention in RGBD images is proposed.In view of the lack of appropriate fusion methods for crossmodal features extracted at different layers by existing deep learning methods of RGBD image salient object detection,this thesis adopted different fusion strategies for crossmodal features at different scales to gradually eliminate irrelevant background regions.We first propose a fusion module is used to locate the location information of the salient region in the upper-level features of the image,Secondly,the gated cross-modal attention module is used to reduce the redundancy information in the single modal features at the lower level of the image and to remove some unnecessary details.Finally,the weighted fusion module is used to fuse the cross-modal features of each feature layer to predict the final salient object detection results.Experimental results show that the proposed RGBD image salient object detection method based on gated cross-modal attention can effectively detect salient objects in complex background images.