一种基于降质学习的低光照图像增强方法  

Low-light Image Enhancement Method Based on Degraded Learning

作　　者：江奎 王中元[4] 黄文心 贾雪梅 王正 胡瑞敏[4] JIANG Kui;WANG Zhongyuan;HUANG Wenxin;JIA Xuemei;WANG Zheng;HU Ruimin(School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China;Hubei Key Laboratory of Intelligent Robot,Wuhan Institute of Technology,Wuhan 430205,China;School of Computer Science and Information,Hubei University,Wuhan 430062,China;School of Computer,Wuhan University,Wuhan 430072,China)

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001 [2]武汉工程大学智能机器人湖北省重点实验室,武汉430205 [3]湖北大学计算机与信息工程学院,武汉430062 [4]武汉大学计算机学院,武汉430072

出　　处：《小型微型计算机系统》2025年第2期353-364,共12页Journal of Chinese Computer Systems

基　　金：国家自然科学基金联合基金项目(U23B2009,U1903214)资助;国家自然科学基金青年项目(6230010538)资助;国家自然科学基金面上项目(62071339,62072347,62171325)资助;智能机器人湖北省重点实验室基金项目(HBIR202311)资助。

摘　　要：低光照图像增强任务旨在提高图像的可见性,同时保持其视觉自然度.针对训练数据缺乏多样性以及恢复图像中细节丢失和颜色失真这两方面问题,基于分布一致性约束,本文提出一种无监督降质学习和数据增广方法用于低光照图像增强,其中包括设计两阶段的网络来学习低光降质的内在特性以及重新恢复低光图像的亮度和纹理细节.受彩色图像成像原理的启发,本文将低光图像增强任务分解为降质学习环境干扰去除,和图像本体细节和颜色细化表达.具体来讲,本文首先从低光输入中估计降质以模拟环境关照因素导致的失真,然后细化内容以恢复漫射导致的内容和对比度损失,并设计一种新颖的降质学习和内容细化网络.在低光图像增强和联合检测任务上的大量实验验证了本文算法的有效性和效率.The low-light image enhancement task aims to improve the visibility of the image while maintaining its visual naturalness.Aiming at the lack of diversity of training data and the loss of details and color distortion in the restored image,based on the distribution consistency constraints,this paper proposes an unsupervised degraded learning and data augmentation scheme for low-light image enhancement.It consists of a two-stage network to learn the intrinsic degradation characteristics and restore the illumination as well as texture details of low-light images.Inspired by the color image formula,this paper decomposes the low-light image enhancement task into the degradation learning and perturbation removal,as well as the content and color refinement.Specifically,this study first estimates the degradation of low-light input to simulate the distortion of ambient lighting color,and then refines the content and contrast to recover the loss of diffuse illumination color.Finally,we propose a novel Degradation Learning and Content Refinement Network(DLCRN).Extensive experiments on low-light image enhancement and joint detection tasks verify the effectiveness and efficiency of our proposed method.

关 键 词：低光图像增强 降质学习 数据增广 编-解码器 扰动去除 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]