Depth Quality-aware RGB-D Salient Object Detection

Salient object detection aim to automatically identify the most attractive objects in the scene by simulating the human visual system,which can help people obtain important information from massive amounts of data and allocate limited computing resources to more important information.In RGB-D salient object detection,the depth picture as auxiliary information plays an important role in the scene where the foreground and background of the RGB picture are not distinguished significantly.Limited by the influence of depth cameras and artificial factors,the quality of depth images varies from scene to scene,and not all features are beneficial to the final detection.Therefore,how to design the network to extract practical features from the bi-stream input information,retain foreground features,shield background features,and achieve the optimal feature complementary fusion state,is particularly important for subsequent detection accuracy.Therefore,this paper proposes two methods to improve the accuracy of salient object detection,and proposes the following solutions:(1)This paper proposes a multi-scale attention module detection network.The bi-stream feature extraction network is used to obtain the preliminarily fused multi-scale RGB-D features,and then the dilated convolution module and the attention module are added layer by layer to complete the task learning.The dilated convolution module in the high-level features expands the receptive field and combines local features with global features,the channel attention module models feature channels,and uses global information to selectively enhance useful information to obtain rich contextual features.The spatial attention module in the low-level features retains local detail information and shields background information.In the end,the network can focus more on foreground objects,and achieve better feature fusion of low-level detailed information under the guidance of high-level semantic information.(2)This paper attempts to integrate a novel depth quality aware subnet into the classic bi-stream structure,aiming to assess the depth quality before conducting the selective RGB-D fusion.Compared with the state-of-the-art bi-stream methods,the major highlight of our method is its ability to lessen the importance of those low-quality,no-contribution,or even negative-contribution depth regions during the RGB-D fusion,achieving a much improved complementary status between RGB and depth.We have devised a novel selective fusion network to make full use of the depth quality aware subnet,achieving a much improved complementary fusion status between RGB and Depth.Using currently widely used evaluation indicators to evaluate the performance of the above methods in multiple datasets,and qualitatively and quantitatively comparative analysis with state-of-the-art methods,it proves that the method proposed in this paper has strong robustness and high detection accuracy.