ELEGAN: AN EFFICIENT LOW LIGHT ENHANCEMENT GAN FOR UNPAIRED SUPERVISION

Low-light image enhancement using deep learning methods in real-world scenarios is challenging, as capturing an enormous paired training dataset is impractical. Many unsupervised low-light image-enhancement methods have recently been explored, although most of them are computationally intense with a complex network structure. This paper proposes a lightweight attention-guided generative adversarial network called ELEGAN for fast low-light image enhancement in a fully unsupervised fashion. We introduce a self-regularized illumination attention map-guided three-layer U-type generator that offers significant speedup by utilizing parallel processing on the encoder side to reduce the inter-dependency between the layers. Additionally, we propose a modified residual dense block for better restoration of features with lesser parameters. We incorporate a global-local discriminator structure in our model. Both visual and quantitative results demonstrate that ELEGAN outperforms state-of-the-art methods in illumination restoration, noise reduction and structure recovery of low-light images.