基于光照可靠性掩膜的低光照图像增强算法  

Low-Light Image Enhancement Based on Illumination Reliability Mask

作　　者：曹晓倩 王旸[1] 刘伟峰 焦登辉 CAO Xiaoqian;WANG Yang;LIU Weifeng;JIAO Denghui(College of Electrical and Control Engineering,Shaanxi University of Science&Technology,Xi’an 710021,China;Huizhou Jufei Optoelectronics Co.,Ltd.,Huizhou,Guangdong 516000,China)

机构地区：[1]陕西科技大学电气与控制工程学院,西安710021 [2]惠州市聚飞光电有限公司,广东惠州516000

出　　处：《计算机工程与应用》2025年第1期263-271,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(62376147);陕西省自然科学基金(2024JC-YBQN-083)。

摘　　要：针对具有局部强光源、局部有色光源的低光照场景图像增强中存在的局部过曝光、局部染色和附加噪声等问题,提出一种基于光照可靠性掩膜的低光照图像增强算法。核心思想是:根据输入图像的光照分量逐像素标记超低光照、过曝光、单通道过曝光等光照不可靠区域并建立“S型”光照可靠性掩膜曲线模型以消除输入低光照图像的局部光照不一致性。以此为基础,设计了由光照估计子模块、光照一致性修正子模块和低光照图像增强子模块组成的基于光照可靠性掩膜的低光照图像增强网络。通过MEF、LIME、DICM、VV等测试数据集的广泛实验表明,所提算法可以显著消除局部过曝光、局部染色及附加噪声等问题,且在NIQE(natural image quality evaluator)与SSIM(structural similarity index)等指标上优于现有算法。相关代码和预训练模型可通过https://github.com/fififft/MLightGAN获取。To address issues of local overexposure,local color cast,and additional noise in low-light image enhancement for scenarios with local strong light sources and colored light sources,a low-light image enhancement algorithm based on illumination reliability masks is proposed.The core idea is marking unreliable illumination regions such as ultra-low illumination,overexposure,and single-channel overexposure pixel-by-pixel according to the illumination components of the input image,followed with an“S-shaped”illumination reliability mask curve model which is established to eliminate local illumination inconsistencies in the input low-light image.Based on this,a low-light image enhancement network is designed,consisting of an illumination estimation sub-module,an illumination consistency correction sub-module,and a low-light image enhancement sub-module.Extensive experiments on test datasets such as MEF,LIME,DICM,and VV demonstrate that the proposed algorithm can significantly eliminate issues of local overexposure,local color cast,and additional noise,and outperforms existing algorithms in metrics such as NIQE(natural image quality evaluator)and SSIM(structural similarity index).The relevant code and pre-trained models can be accessed at https://github.com/fififft/MLightGAN.

关 键 词：低光照图像增强 局部光照不一致 光照可靠性掩膜 光照估计 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]