Dual-Branch Infrared and Visible Image Fusion Framework

This paper introduces an innovative infrared and visible image fusion network that leverages reversible neural networks and self-attention mechanisms within a dual-branch feature extraction structure. The reversible neural network focuses on extracting high-frequency local features, while the self-attention mechanism, integrated into the Stoken Transformer (ST) module, efficiently captures low-frequency global dependencies through a novel Fold and Unfold operation. By decomposing traditional global attention into the product of sparse association mapping and low-dimensional attention, the ST module effectively enhances the representation of low-frequency global features. Experimental evaluations demonstrate that the proposed approach outperforms state-of-the-art (SOTA) methods, achieving superior performance in downstream tasks such as object detection and semantic segmentation. Notably, it achieves 80.98 [email protected] on the M3FD detection task, highlighting its effectiveness in advancing infrared and visible image fusion technology.