An Improved Low-Light Image Enhancement Algorithm Based on Deep Learning

At present, most low-light image enhancement algorithms rely on artificial design with a priori information and constraints, but it cannot accurately capture deep structural features of the images. In this paper, we propose an improved algorithm with EnlightenGAN as the base framework and consider the phenomenon that the distance reflected by Mean Square Error (MSE) is quite different from human intuition. By introducing Structural Similarity (SSIM) in the loss function, we try to improve the structural similarity of the enhanced images to make them more consistent with natural and human intuition. For the instability problem during Generative Adversarial Networks (GAN) training, the relatively loose and easy-to-compute Energy-Based GAN (EBGAN) is used instead of Wasserstein GAN (WGAN). The improved algorithm is finally tested by using Low-light Image Enhancement (LIME) dataset as well as the test dataset of EnlightenGAN. The Peak Signal to Noise Ratio (PSNR) and SSIM values of the images processed by using the modified algorithm are calculated and compared with other algorithms, and the results show the proposed method is effective.