RGB-T salient object detection based on progressive alignment and multiscale fusion
Fucai Zhang · 2025
Existing RGB-T salient object detection methods enhance the detection performance by fusing thermal infrared (TIR) images. However, there are significant deficiencies in the design of the cross-modal information fusion mechanism: the modal features are not fully aligned, resulting in feature offset, insufficient cross-layer interaction, and interference from redundant background noise. To address this, a progressive alignment and multi-scale fusion method is proposed. This method designs a novel three-stage alignment framework at the feature level, pixel level, and semantic level, and introduces a cross-modal attention mechanism to achieve fine alignment. Secondly, a multi-scale convolution and dense transmission strategy is adopted for feature fusion, and a feature denoising module is used to effectively suppress background noise. Finally, through the decoding operations of concatenation and dimensionality reduction, an accurate prediction map of the salient object is generated.