CMCLNet Cross-Modality Attention Fusion and Cross-Level Feature Interaction for RGBD salient object detection
Lifang Xiao, Huimei Chen, Qingzhen Xu, Qiang Chen · 2023
So as to address the issue of edge defects in RGB-D based salient object detection, we propose a novel structure called Cross Modality Attention Fusion and Cross-Level Feature Interaction (CMCL), with the aim of improving edge quality and effectively extracting salient object from complex backgrounds. Our network is an encoder-decoder structure, specifically, (1) a cross-modality attention fusion (CMAF) module is proposed, which enhance and fuse relevant information of RGB images and depth map, then output enhanced features; (2) the cross-level feature interaction (CLFI) module is proposed to improve edge quality in the encoder stage, combining low-level edge features with high-level semantic information; (3) In the decoder stage, we put forward a gate fusion unit (GU) to reduce background noise, enhancing the significant features of the images. We carry on experiments quantitatively and qualitatively on six publicly available RGB-D datasets and compared them using four evaluation indicators. The results show that our network outperforms mainstream salient detection methods to a certain extent, which is of great significance for solving the key problem of salient object detection in computer vision.