Dual-Stream Bilateral Feature Fusion Network for RGB-D Salient Object Detection

To address the challenges associated with unstable depth map quality in RGB-D salient object detection, this paper introduces a dual-stream network (DSN) architecture that independently processes RGB and depth image inputs while merging these modalities using a multi-scale fusion strategy. Depth maps, however, are often affected by noise and incomplete regions due to sensor limitations, resulting in quality fluctuations that hinder effective multi-modal feature fusion. To mitigate this issue, we incorporate gated convolution into the depth stream, enabling dynamic adjustment of the feature extraction process based on depth map quality. This approach effectively suppresses low-quality depth information, facilitating the efficient integration of depth and RGB features. Consequently, the network can extract valuable geometric information from low-quality depth maps, maintaining high detection accuracy in complex scenes. Extensive experiments on multiple public datasets show that the proposed method significantly enhances performance in salient object detection tasks.