TMFBSOI: A Three-Stream Multi-Feature Fusion RGBD Salient Object Detection Network Based on Specific Object Imaging
R. Zhang, Jun Li, Shuqi Luo, Weizhao Chen, Yaojin Xie · 2024
The effectiveness of depth information in salient object detection has long been proven .However, how to apply depth information is still under exploration.Despite the notable achievements of recent deep learning approaches for RGBD salient object detection, they have failed to fully integrate original and accurate information to express the details of salient objects in RGBD images, leading to blurred or incomplete detection results.Therefore, a three-stream multi-feature fusion model based on specific object imaging fusion, TMFBSOI, is proposed.The three-stream structure ensures the acquisition of original RGBD image information, fully retaining the relevant features existing in the RGB and depth maps; unlike general models that integrate attention mechanisms in hierarchical features; the object imaging method directly obtains the three-dimensional salient target area in the original image through simultaneous imaging of the RGB map and depth map, which is a complete manifestation of the attention mechanism, fully preserving the details and accuracy of RGBD salient target features.Our empirical evaluations demonstrate that TMFBSOI outperforms other models of comparable complexity when trained on a small dataset across all standard metrics.Moreover, in challenging scenes, such as occlusions or low-light conditions, the proposed model also shows excellent performance.