A Modern Physical GAN for Unsupervised Low-Light Image Enhancement via Illumination Estimation

• Main Authors: Liu, Tong, Xu, WenDa, Liu, YuFeng, Liu, SiYuan, Chen, XiaoLu
• Format: Conference Proceeding
• Language: English
• Subjects: Convolutional Neural Network; Generative Adversarial Network; Generative adversarial networks; Generators; Illumination Estimation; Image resolution; Lighting; Low-Light Image Enhancement; Retinex Model; Supervised learning; Surveillance; Visualization
• ISSN: 2161-2927
• DOI: 10.23919/CCC58697.2023.10240033

Low-light environments can cause severe degradation in the visual quality of captured images, leading to the failure of advanced visual perception algorithms and creating hazards in areas such as unmanned vehicles, security surveillance, and visual localization. While deep learning has driven the advancement of Low-Light Image Enhancement (LLIE) techniques, the generalization capability of supervised learning methods is limited by the quality and size of paired datasets. To address this challenge, this paper proposes a novel unsupervised LLIE method that combines a physical model with a Generative Adversarial Network (GAN). The proposed method constructs a lightweight, fully convolutional GAN network that supports arbitrary resolution inputs and is based on modern convolutional blocks. Unlike the end-to-end approach, the generator takes low-light images and illumination as inputs, and combines them with physical models to generate enhanced results. Additionally, a multi-scale deep supervision mechanism is introduced in the discriminator to improve the visual quality of the generated images. The proposed method is compared with existing mainstream methods both qualitatively and quantitatively, demonstrating its lightweight, effectiveness, and superiority.