LIENet: A low-light image enhancement network for extreme darkness

The poor visibility of images acquired in low-light environments has seriously affected the development of computer vision-related fields. Most previous work on low-light enhancement either only focuses on a single task, or only learns a single-channel mapping between input low-light images and real images. These methods rely on joint light intensity adjustments to pairs of long- and short-exposure images from a specific camera model, which is insufficient to handle low-light images in special imaging environments. In this paper, we propose a light enhancement framework that jointly adjusts light intensity, image denoising, and super-resolution enhancement. We design a low-light image enhancement network (LIENet) from three aspects of the overall framework, network structure, and loss function. Specifically, a Perceptual Detail Generative Adversarial Network (PD-GAN) is first designed to enhance texture and suppress noise. Second, the encoder-decoder structure is used to restore the illumination, and the residual dense block finely controls the noise characteristics. Then, a new super-resolution enhancement network structure is designed to avoid the blurring phenomenon in the enhanced image. Finally, fractional calculus is used to extract noise and weak contour information, and a loss function is designed to improve texture and suppress noise. Comprehensive experiments on publicly available datasets show that the proposed LIENet significantly outperforms the state-of-the-art methods in terms of visual performance.