MFENet: RGB-d salient object detection network with multiscale feature enhancement
Shanshan Wang, Bin Li, Wenhua Cui · 2025
This paper proposes a multi-scale feature enhanced RGB-D salient object detection network (MFENet). It consists of a multi-branch flow network, namely the RGB image and depth image feature learning network, as well as the cross-modal shared learning network. In the two image feature learning networks, this paper constructs a granularity feature attention module (GFA). It employs the Otsu algorithm to divide the depth image into multiple local regions, in order to fully exploit the spatial prior information of the depth image and thereby enhance the accuracy of salient object detection. In the shared learning network, this paper constructs a Feature Pyramid Module (FPM). It adopts various sampling rates of dilated convolution to explore complementary information of RGB images and depth images, in order to enhance the shared feature outputs. This paper validates the effectiveness of the MFENet model on four significant benchmark datasets for object detection. The highest accuracy rate of its significance detection reached 94.8%, and the average accuracy rate was 93.2%. The experimental results demonstrate that MFENet significantly enhances the robustness and accuracy of the saliency object detection model in complex visual scenarios.