Alignment-free RGB-T salient object detection: a semantic-guided deformable alignment and asymmetric fusion network

RGB-T salient object detection aims to segment salient objects by exploiting complementary information from visible and thermal infrared modalities. However, existing methods often rely on manually aligned pairs and struggle with large-range offsets, scale variations, and local nonlinear deformations during acquisition. These discrepancies may weaken cross-modal correlations and introduce unsuitable fusion cues. To address these challenges, we propose a Semantic-guided Deformable Alignment and Asymmetric Fusion Network (SDAF-Net) for alignment-free RGB-T salient object detection. First, a multi-scale semantic constraint module extracts a shared saliency-aware representation and generates semantic priors for deformable feature sampling and fusion. Then, a semantic-guided dense deformation alignment module estimates a task-optimized Cross-modal Deformation Offset Map through coarse-to-fine refinement. The offset map resamples multi-scale thermal features on the RGB reference feature grid and provides spatially adaptive cross-modal cues. Finally, an asymmetric rectification fusion module uses task-driven masks generated jointly from RGB and resampled thermal features to calibrate RGB responses while preserving the original thermal features for fusion. Experiments on unaligned, weakly aligned, and aligned RGB-T benchmarks demonstrate competitive performance and more complete saliency maps in complex scenarios.