MDFN:Multi-scale Dense Fusion Network for RGB-D Salient Object Detection
Xinyu Wen, Feng Wang, Zhengyong Feng, Jun Pin Lin, Chengkun Shi · 2023
In RGB-D salient object detection, the low quality of the salient maps produced after detection is caused by the unclear contents of some RGB-D images and the poor utilization of image information by some methods. To solve the above problems, in this paper, we propose a Multi-scale Dense Fusion Network (MDFN) to improve RGB-D image understanding and enhance the performance of salient object detection. First, in the design of the encoder, we chose the Pyramid Pooling Transformer (P2T) as the backbone network of the encoder, and take advantage of the strong image understanding capability of the P2T to prevent the loss of important information when extracting RGB-D features. For the limitations such as less depth map information and noise interference, we propose the Depth Attention Enhancement Module (DAEM) to enhance the depth feature representation, and design the RGB-Depth Fusion Module (RDFM) to achieve RGB-D information matching. In the decoder design, to improve the utilization of image information, we propose the Dual-dense Fusion Module (DFM) to fully integrate the features of each layer and achieve multiple applications of the object information. Compared with 7 SOTA methods on 3 public datasets, MDFN achieves excellent performance on 4 evaluation metrics $(S_{\alpha}, F_{\beta}, E_{\xi}, M)$ and saliency maps, which proves the effectiveness of our method.