An Improved CycleGAN-Based Model for Low-Light Image Enhancement

Abstract

The low-light image enhancement is a challenging and hot research issue in the image processing field. In order to enhance the quality of low-light images to obtain full structure and details, many low-light image enhancement algorithms have been proposed and deep learning-based methods have achieved great success in this field. However, most of the deep learning methods require paired training data, which is difficult to obtain. And the overall visual quality of the enhanced image is still not very satisfying. To deal with these problems, an unsupervised low-light image enhancement model based on an improved Cycle-Consistent Generative Adversarial Networks (CycleGAN) is proposed in this paper. In the proposed model, a low-light enhancement generator of the CycleGAN network is constructed based on an improved U-Net structure, and the adaptive instance normalization (AdaIN) is designed to learn the style of the normal light image. In particular, a detail enhancement method based on multi-layer guided filtering is added to the proposed model, which can improve the quality and visual pleasantness of image enhancement. In addition, a joint training strategy based on structural similarity is presented, to strengthen the constraints on generating more realistic and natural images. At last, extensive experiments are conducted and the results show that the proposed method can accomplish the task of transferring low-light images to normal light and outperform the state-of-the-art approaches in various metrics of visual quality.