Residual-Unet Generative Flow Model for Low-Light Image Enhancement

Low-light image enhancement (LLIE) task based on deep learning learns the mapping from low-light image to normal exposure image by using supervised or unsupervised training. In the majority of current methods, the utilization of Convolutional Neural Networks (CNNs) substantially constrains the network's capacity to assimilate comprehensive global information. Simultaneously, the reconstruction loss or adversarial loss commonly employed tends to inadequately gauge the visual disparity between predictions and targets, resulting in the emergence of blurred regions in the enhanced results. In this paper, we therefore propose a Residual-Unet Generative Flow Model for Low-light Image Enhancement (RUGF), which consist of a feature extraction module and a flow learning based reconstruction module. We first propose a residual-based Unet encoder during deep feature encoding in the feature extraction module, which can compensate for the shortcomings in global information processing and has advantages in processing depth information. In addition, our network adopts a negative logarithmic likelihood loss function, which can better constrain the model to learn the complex distribution of the reference image, thereby avoiding the local area blurring that exists in existing methods. We conducted extensive testing on two benchmark datasets, and the actual results confirmed the superiority and effectiveness of our proposed method.