An unsupervised fusion method for infrared and visible image under low-light condition based on Generative Adversarial Networks

The aim of fusing infrared and visible images is to achieve high-quality images by enhancing textural details and obtaining complementary benefits. However, the existing methods for fusing infrared and visible images are suitable only normal lighting scenes. The details of the visible image under low-light conditions are not discernible. Achieving complementarity between the image contours and textural details is challenging between the infrared image and the visible image. With the intention of addressing the challenge of poor quality of infrared and visible light fusion images under low light conditions, a novel unsupervised fusion method for infrared and visible image under low_light condition (referred to as UFIVL) is presented in this paper. Specifically, the proposed method effectively enhances the low-light regions of visible light images while reducing noise. To incorporate style features of the image into the reconstruction of content features, a sparse-connection dense structure is designed. An adaptive contrast-limited histogram equalization loss function is introduced to improve contrast and brightness in the fused image. The joint gradient loss is proposed to extract clearer texture features under low-light conditions. This end-to-end method generates fused images with enhanced contrast and rich details. Furthermore, considering the issues in existing public datasets, a dataset for individuals and objects in low-light conditions (LLHO https://github.com/alex551781/LLHO) is proposed. On the ground of the experimental results, we can conclude that the proposed method generates fusion images with higher subjective and objective quantification scores on both the LLVIP public dataset and the LLHO self-built dataset. Additionally, we apply the fusion images generated by UFIVL method to the advanced computer vision task of target detection, resulting in a significant improvement in detection performance.