基于轻量化神经网络的低照度图像增强算法  

A Low-Light Image Enhancement Algorithm Based on Lightweight Neural Network

作　　者：刘芸萌 龙永红[1] 李欣 LIU Yunmeng;LONG Yonghong;LI Xin(College of Rail Transit,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学轨道交通学院,湖南株洲412007

出　　处：《湖南工业大学学报》2025年第4期41-47,共7页Journal of Hunan University of Technology

基　　金：湖南省自然科学基金资助项目(2024JJ7144)。

摘　　要：针对低照度场景下图像亮度低的问题,提出了一种基于轻量化神经网络的无监督低照度图像增强算法。提出一种可学习的内容自适应S型亮度映射曲线,不仅能扩大亮度调整范围,且能在保证亮度的情况下保持良好的对比度;此外,设计了一个轻量化的亮度曲线估计网络,网络采用无监督训练,通过学习输入图像与拟合曲线之间的映射关系,解决了标签数据获取难的问题。实验结果表明,所提轻量化图像增强网络的计算量较小,有效减少了计算时间,在不同数据集上均取得了良好表现,为低照度场景下的图像增强提供了有效的解决方案。In view of the low image brightness in low-light scenarios,an unsupervised low-light image enhancement algorithm has thus been proposed based on a lightweight neural network.A learnable content-adaptive S-shaped brightness mapping curve is introduced to expand the brightness adjustment range and maintain good contrast while ensuring brightness.A lightweight brightness curve estimation network is designed,which adopts unsupervised training to learn the mapping relationship between the input image and the fitted curve,thus solving the problem of difficult access to labeled data.Experimental results show that the proposed lightweight image enhancement network is characterized with a low computational cost so as to effectively reduce computation time,and achieve a good performance on different datasets,thus providing an effective solution for image enhancement in low light scenarios.

关 键 词：S型亮度映射曲线 轻量化神经网络 无监督学习 图像增强 

分 类 号：TP391.413[自动化与计算机技术—计算机应用技术]