Dynamic Frequency Separated Fusion Network for RGB-T Semantic Segmentation

RGB-T semantic segmentation fuses visible (RGB) and thermal infrared (TIR) images to enhance model robustness under complex lighting conditions. However, most existing methods employ unified fusion strategies in the spatial domain, limiting their ability to fully exploit the complementary characteristics of the two modalities across different frequency components. To address this issue, we propose a novel fusion model that achieves finer cross-modal collaboration through frequency-aware feature disentanglement and adaptive fusion mechanisms. Specifically, the model first enhances the multi-granularity representation of dual-modality features using attention mechanisms. The features are then transformed into the frequency domain, where a dynamic band disentanglement mechanism adaptively separates high-frequency details from low-frequency semantics based on the input content. A heterogeneous fusion pathway is further designed to differentially aggregate high- and low-frequency components, emphasizing cross-modal detail alignment in the high-frequency path and semantic consistency in the low-frequency path. The proposed model is evaluated on three widely-used RGB-T segmentation benchmarks, namely MFNet, PST900, and FMB, and achieves state-of-the-art performance, demonstrating its effectiveness and superiority.