Infrared and visible image fusion based on asymmetric coupling networks

Traditional infrared and visible image fusion methods rely on well-designed decomposition and fusion frameworks. Similarly, deep learning-based fusion networks often involve well-designed fusion methods. Although both methods have achieved good integration performance, they both have a lot of traces of human intervention. Therefore, this paper proposes a self-supervised fusion method based on asymmetrically coupled networks. It does not rely on supervised data for network training, nor does it need to design fusion methods to achieve end-to-end fusion manually. The technique consists of two asymmetric subnetworks, each focusing on a different source image feature. Comparative experiments on TNO, roadscene, and M3FD datasets show that the fusion method can better preserve the contour and texture details of the source image, and save more computing resources.