Unsupervised learning based dual-branch fusion low-light image enhancement

Distortion-free enhancement on images captured under low-light conditions has always been a challenging problem in computer vision. Although many image enhancement methods have been proposed, most existing algorithms cause artifacts and amplify the noise in the enhanced image, which greatly affect the visual perception of the enhanced image. To address these problems, this paper proposes an unsupervised learning based dual-branch fusion low-light image enhancement algorithm, which can learn the map way of low-light images to normal-light images from unpaired low-light and normal-light datasets. The network is consisted of dual branches, the upper branch is a refinement branch focusing on noise suppression, and the lower branch is a U-Net-like global reconstruction branch based on the attention mechanism for high-quality image generation. The discrimination network adopts the multi-scale discrimination structure of feature pyramid to enhance the global consistency and avoid local overexposure. The loss function is also improved, and a new fidelity cycle consistency loss is introduced to further improve the quality of image texture information recovery. Qualitative and quantitative experimental results show that the proposed method can effectively suppress the generation of artifacts and noise amplification of enhanced images.