Swin transformer and ResNet based deep networks for low-light image enhancement

Low-light image enhancement is a long-term low-level vision problem, which aims to improve the visual quality of images captured in low illumination environment. Convolutional neural network (CNN) is the foundation of the majority of low-light image enhancement algorithms now. The limitations of CNN receptive field lead to the inability to establish long-range context interaction. In recent years, Transformer has received increasing attention in computer vision due to its global attention. In this paper, we design the Swin Transformer and ResNet-based Generative Adversarial Network (STRN) for low-light image enhancement by combining the advantages of ResNet and the Swin Transformer. The STRN consists of a U-shaped generator and multiscale discriminators. The generator is composed of a shallow feature extraction, a deep feature extraction, and an image reconstruction module. To calculate the global and local attention, we alternately use Swin Transformer blocks and ResNet in the deep feature processing module. The self perceptual loss and the spatial consistency loss are employed to constrain the random paired training of STRN. The experimental results on benchmark datasets and real-world low-light images demonstrate that the proposed STRN achieves state-of-the-art performance on low-light image enhancement tasks in terms of visual quality and evaluation metrics.