SRFusion An infrared and visible images fusion methodbased on structural reparameterization

Abstract Fusion of infrared and visible images amalgamates their respective advantages while circumventingtheir drawbacks, thereby retaining the intricate details present in the visible spectrum and augment-ing the thermal radiation information inherent in the infrared domain. However, few studies havefocused on the balance between algorithm speed and fused image quality in complex lighting condi-tions, which is critical for downstream tasks. To this end, we propose An infrared and visible imagesfusion method based on structural reparameterization, named SRFusion. The method innovativelyintegrates a Structural Reparameterization Module (SRM), adeptly extracting and processing featuresfrom the input images with exceptional efficiency. The methodology bifurcates the extracted featuresinto distinct components, which are then recombined through a sequence of convolutional operationsto produce a fused image. Notably, the SRM operates with a multi-branch architecture during thetraining phase, optimizing performance, and transitions to a streamlined, branch-free topology forinference, enhancing computational efficiency. A comparative analysis with the most advanced fusionmethods has demonstrated that SRFusion not only achieves the SOTA performance on most metricsof fusion image, but also delivers a more intricate level of detail and a better visual perception for human.