Low-Light Image Enhancement via Global-Local Collaborative Transformer

Low-light image enhancement aims to refine the lighting conditions of low-light images. However, existing enhancement methods lack an explicit learning mechanism for perceiving exposure, making it difficult to handle unevenly exposed low-light images. To address this issue, we propose a global-local collaborative transformer, an enhancer inspired by image editing experts, which exploits the collaboration between global and local adjustment to adaptively enhance low-light images with complex exposure. Specifically, the proposed model independently learns local and global self-attention to implement the heterogeneous adjustment required for low-light regions with different exposure, and then employs an illumination difference map to dynamically determine the contributions of the two types of adjustment. Meanwhile, we develop a feed-forward neural network integrating globality and locality to further promote the efficacy of the proposed collaborative learning paradigm. In addition, a history-aware contrastive loss is designed to encourage the model to generate high-quality enhanced images that approximate the distribution domain of the normal-light images. Experiments performed on several popular public datasets demonstrate that the proposed enhancer outperforms state-of-the-art models. In particular, it shows competitive performance for low-light images with uneven exposure. Our code is available at: https://github.com/Shecyy/GLCFormer.