Hierarchical interaction fusion network for RGB-D salient object detection

RGB-D salient object detection aims to exploit the spatial and geometric cues provided by depth maps to complement RGB images in complex scenes. However, most existing methods adopt static fusion strategies, overlooking the hierarchical representation characteristics across different network stages. To address these issues, we propose a Hierarchical Interaction Fusion Network (HIFNet) based on a three-stream encoder–decoder architecture. HIFNet is designed to explicitly model cross-modal complementary information and preserve modality-specific distinctive features. Specifically, in the encoder stage, a bi-directional attention gated fusion (BAGF) module is introduced for shallow features. This module facilitates inter-modal interaction via bidirectional cross-modal attention and employs a spatial gating mechanism to adaptively integrate the enhanced features, effectively suppressing low-level noise. For deep features, a dual-branch complementary fusion (DCF) module is designed to capture shared representations and long-range semantic dependencies via parallel branches, thereby enhancing global saliency perception. In the decoder stage, an attention refinement decoder (ARD) is designed to progressively refine predictions by integrating high-level semantics with low-level spatial details. Extensive experiments on multiple RGB-D SOD benchmarks demonstrate that HIFNet achieves highly competitive performance compared with state-of-the-art methods across four evaluation metrics.