Research On RGB-D Salient Object Detection Based On Cross-modality And Multi-level Fusion

Salient object detection(SOD)is dedicated to simulating the human visual system and capturing the most eye-catching objects in a given scene.The deep-learning-based SOD methods have shown the great value of researching and have been applied in many aspects in real lives.In recent years,depth cameras have become popular in mobile devices,which makes the acquisition of depth images more convenient.The distance information carried by the depth images can provide geometric feature information of targets in the scene.The performance of RGB(red-green-blue)-based salient object detection can also be improved by combining depth(D)clues.In RGB-D(red-green-blue-depth)salient object detection task,how to fuse the cross-modality features between RGB and depth(D)and how fully utilize the multi-level features are two important research.For the fusion of cross-modality features,many researchers ne-glect the differences between these two types of information,which ultimately leads to the inefficient fusion of the cross-modality features.As researchers focus more on the fusion of cross-modality features,many works fail to take full use of the multi-level features.This paper focus on the fusion of cross-modality features and multi-level features,the main inno-vations are as follows:Contribution measure module(CMM)is proposed to constrain cross-modality features according to the contribution to the generation of the saliency map,based on this idea,this paper further suggests the cross-Modal features fusion module(CMFM)?Based on the discovery that semantic information is distributed in different channels,this paper proposes the channel-wise guided fusion module(CGFM)to efficiently promote the communication of information among multi-level features and allow more complementary information to be fused?To achieve better performance,a subnet is proposed to enhance the fusion of features.These three innovations constitute the cross-modality and multi-level feature fusion net-work(CMFF-Net)proposed in this paper.Extensive experiments show that CMFF-Net out-performs 16 state-of-the-art methods.Specifically,compared with the second-best methods on NJU2K,our S_α,MAE,E_γand F_βvalues are higher by 1.3%,0.3%,0.8%,and 1.2%respectively,and the results on NLPR,SIP,STERE,DUT-RGBD are also in the leading place.