Infrared and Visible Image Fusion Algorithm Based on Cross-Modal Attention Mechanism

Abstract

As a critical branch of multi-modal image processing, infrared and visible image fusion boasts high application value in intelligent security and has drawn widespread global research attention. Enhancing model feature extraction is a core scientific challenge in this field. This paper presents a dual-branch infrared and visible image fusion algorithm based on feature decomposition. In fusion tasks, shared features characterize global information while private features focus on local details. To boost feature representation, we design a feature decomposition module that splits shallow features into shared and private components: a coarse-grained branch with medium-to-large receptive fields handles global shared features, and a fine-grained branch with small receptive fields extracts local private features. Parallel dual-branch processing of decomposed features enables precise data structure mining, reduces redundancy, and efficiently captures key information. Experiments on four mainstream public datasets validate that the proposed algorithm surpasses state-of-the-art methods in information extraction, detail preservation and fusion performance.