DGSNet: An RGB-D Salient Object Detection Network Based on a Deep Guidance Strategy

The existing RGB-D salient object detection (SOD) networks utilize multilevel RGB and depth features to detect salient objects. However, most models do not fully exploit the distinct advantages provided by features located at different levels and lack mechanisms for regulating the impact of depth information. This paper introduces a novel RGB-D SOD network based on a deep guidance strategy that contains two branches. The deep features extracted by one branch serve as guidance for the other branch, which produces the final detection results. The network incorporates a novel self-attention-weighted fusion mechanism, which significantly enhances the robustness exhibited in scenarios with low-quality depth maps. Additionally, a global enhancement module (GEM) for handling dense contextual information efficiently utilizes rich global data. The experimental results obtained on six commonly used benchmark datasets demonstrate that the proposed network model outperforms several existing models across five widely recognized evaluation metrics.