Unsupervised Low-Light Image Enhancement Based on Explicit Denoising and Knowledge Distillation  

作　　者：Wenkai Zhang Hao Zhang Xianming Liu Xiaoyu Guo Xinzhe Wang Shuiwang Li 

机构地区：[1]College of Computer Science and Engineering,Guilin University of Technology,Guilin,541006,China [2]Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technology,Guilin,541004,China

出　　处：《Computers, Materials & Continua》2025年第2期2537-2554,共18页计算机、材料和连续体(英文)

基　　金：support by the Guangxi Natural Science Foundation(Grant No.2024GXNSFAA010484);the NationalNatural Science Foundation of China(No.62466013),this work has been made possible.

摘　　要：Under low-illumination conditions, the quality of image signals deteriorates significantly, typically characterized by a peak signal-to-noise ratio (PSNR) below 10 dB, which severely limits the usability of the images. Supervised methods, which utilize paired high-low light images as training sets, can enhance the PSNR to around 20 dB, significantly improving image quality. However, such data is challenging to obtain. In recent years, unsupervised low-light image enhancement (LIE) methods based on the Retinex framework have been proposed, but they generally lag behind supervised methods by 5–10 dB in performance. In this paper, we introduce the Denoising-Distilled Retine (DDR) method, an unsupervised approach that integrates denoising priors into a Retinex-based training framework. By explicitly incorporating denoising, the DDR method effectively addresses the challenges of noise and artifacts in low-light images, thereby enhancing the performance of the Retinex framework. The model achieved a PSNR of 19.82 dB on the LOL dataset, which is comparable to the performance of supervised methods. Furthermore, by applying knowledge distillation, the DDR method optimizes the model for real-time processing of low-light images, achieving a processing speed of 199.7 fps without incurring additional computational costs. While the DDR method has demonstrated superior performance in terms of image quality and processing speed, there is still room for improvement in terms of robustness across different color spaces and under highly resource-constrained conditions. Future research will focus on enhancing the model’s generalizability and adaptability to address these challenges. Our rigorous testing on public datasets further substantiates the DDR method’s state-of-the-art performance in both image quality and processing speed.

关 键 词：Deep learning low-light image enhancement real-time processing knowledge distillation 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]