MPCANet: Multi-physical prior guided cross-modal attention and fusion network for RGB-T salient object detection

RGB-T salient object detection (SOD) fuses the fine details of RGB with the robustness of thermal infrared (TIR) to segment salient objects across diverse scenes and lighting conditions. However, existing approaches face three main challenges: (i) thermal noise and registration errors contaminate attention at early fusion, (ii) cross-modal misalignment and detail preservation are difficult to balance, and (iii) thermophysical properties (e.g., heat diffusion and inertia) are under-modeled, leading to instability in low-contrast and dynamic settings. Thermophysical priors could provide interpretable cues on boundaries, diffusion, inertia, and material differences, suppressing noise at the source and strengthening structural reliability. When combined with high-level semantics that provide global task relevance and region-level attention, interaction is reinforced only in semantically relevant and physically reliable areas. Building on this principle, we propose a semantics-physics integrated cross-modal fusion framework. The key idea is to inject thermophysical priors into attention logits before normalization to suppress upstream thermal noise, while semantic saliency dynamically guides the fusion strategy by controlling the receptive field and sampling for bi-directional alignment, thereby achieving a more robust and precise feature fusion that balances global consistency with local sharpness. During the decoding process, a boundary-semantic coupling restores fine structures, and a Training Objective jointly supervises pixels, regions, boundaries, and features to prevent internal inconsistency. Extensive experiments on public benchmarks demonstrate state-of-the-art performance with sharper boundaries and more coherent regions. Code is available at https://github.com/taromm/MPCANet.