Lowlight image enhancement based on unsupervised learning global-local feature modeling

Abstract

Inspired by the generative adversarial network EnlightenGAN, we propose a novel low-light image enhancement algorithm based on unsupervised learning global-local feature modeling (GLFE). The algorithm has two stages: generation and discrimination, including global and local feature modeling network, and global and local discriminator. First of all, Swin-Transformer Block is innovatively introduced in the global feature modeling of the generation stage. Its shift window mechanism can conduct long-distance feature dependence modeling of the input image with less memory consumption, and well extract the features of image color, texture and shape, so as to effectively suppress noise and artifacts. Secondly, in the local feature modeling, the U-net branch based on grayscale spatial attention guidance can well capture the detailed information such as image edges and corner points. In the discrimination stage, deep and shallow feature fusion modules are added to enhance the discrimination ability, and the inconsistency is suppressed by learning the spatial filter contradiction information, so that the shallow representation information and deep semantic information guide each other, and the reasoning is almost no overhead, so that the enhanced image has uniform illumination intensity. Thanks to the synergistic effect of the above three innovative aspects, GLFE can achieve greater performance improvement compared with EnlightenGAN. Compared with the existing low-light enhancement algorithms, the algorithm achieves SOTA level performance in several public datasets.