The purpose of infrared and visible image fusion is to integrate the complementary information from heterogeneous images in order to enhance their detailed scene information. However, existing deep learning fusion methods suffer from an imbalance between fusion performance and computational resource consumption. Additionally, fusion layers or fusion rules fail to effectively combine heteromodal feature information. To address these challenges, this paper presents a novel algorithm called infrared and visible image fusion network base on fast edge convolution (FECFusion). During the training phase, the proposed algorithm enhances the extraction of texture features in the source image through the utilization of structural re-parameterization edge convolution (RECB) with embedded edge operators. Subsequently, the attention fusion module (AFM) is employed to sufficiently fuze both unique and public information from the heteromodal features. In the inference stage, we further optimize the training network using the structural reparameterization technique, resulting in a VGG-like network architecture. This optimization improves the fusion speed while maintaining the fusion performance. To evaluate the performance of the proposed FECFusion algorithm, qualitative and quantitative experiments are conducted. Seven advanced fusion algorithms are compared using MSRS, TNO, and M3FD datasets. The results demonstrate that the fusion algorithm presented in this paper achieves superior performance in multiple evaluation metrics, while consuming fewer computational resources. Consequently, the proposed algorithm yields better visual results and provides richer scene detail information.
Keywords: deep learning; edge operator; image fusion; infrared and visible images; structural re-parameterization.