Low-light image enhancement based on U-Net and fusion strategy

Abstract

Normal light images are a prerequisite for many visual tasks, making single image enhancement crucial in computer vision. While physical model-based approaches have demonstrated promising results in this domain at an early stage, they often produce undesired anomalies in their outputs, making it difficult to apply them universally across all scenarios. Conversely, learning-based methods yield more authentic outcomes. However, the enhancement capabilities of machine learning algorithms are limited by the lack of pairs of normal light images, and many algorithms focus on a single enhancement task, ignoring the role of high-level tasks in guiding enhancement. This paper introduces a low-light image enhancement algorithm rooted in a fusion strategy that consists of two subtasks, visibility restoration and realism improvement tasks, and combines the advantages of physical modeling and learning methods. In the visibility restoration stage, the fusion of features from upstream and downstream tasks enables our network to take full advantage of the guidance of the upstream task to the downstream task to recover results closer to nature; in the fidelity enhancement stage, the Retinex enhancer and the refinement enhancement network are introduced to enhance the fidelity of the image. Finally, the results of the two subtasks are fused to obtain the output. Experiments substantiate that the algorithm excels in terms of visual effects and objective metrics across both synthetic data and real-life scenes.