MGRCFusion: An infrared and visible image fusion network based on multi-scale group residual convolution

Abstract

The purpose of fusing infrared and visible images is to obtain an informative image that contains bright thermal targets and rich visible texture details. However, the existing deep learning-based algorithms generally neglect finer deep-level multi-scale features, and only the last layer of features is injected into the feature fusion strategy. To this end, we propose an optimized network model for deeper-level multi-scale features extraction based on multi-scale group residual convolution. Meanwhile, a dense connection module is designed to adequately integrate these multi-scale feature information. We contrast our method with advanced deep learning-based algorithms on multiple datasets. Extensive qualitative and quantitative experiments reveal that our method surpasses the existing fusion methods. Furthermore, ablation experiments illustrate the excellence of the multi-scale group residual convolution module for infrared and visible image fusion.