基于渐进式双网络模型的低曝光图像增强方法  

A Low-Exposure Image Enhancement Based on Progressive Dual Network Model

作　　者：黄淑英 胡威 杨勇[2] 李红霞 汪斌 HUANG Shu-Ying;HU Wei;YANG Yong;LI Hong-Xia;WANG Bin(School of Computer Science and Technology,Tiangong University,Tianjin 300387;School of Information Management,Jiangxi University of Finance and Economics,Nanchang 330032;School of Soft-ware and Internet of Things Engineering,Jiangxi University of Finance and Economics,Nanchang 330032)

机构地区：[1]天津工业大学计算机科学与技术学院,天津300387 [2]江西财经大学信息管理学院,南昌330032 [3]江西财经大学软件与物联网工程学院,南昌330032

出　　处：《计算机学报》2021年第2期384-394,共11页Chinese Journal of Computers

基　　金：国家自然科学基金(61862030,61662026);江西省自然科学基金(20182BCB22006,20181BAB202010,20192ACB20002,20192ACBL21008)资助。

摘　　要：传统的图像增强方法对低曝光图像进行增强时,通常只考虑到了亮度的提升,忽略了增强过程中带来的噪声放大问题.而当前基于深度学习的方法利用端到端的网络直接学习低曝光图像到正常图像的映射关系,忽略了低曝光图像形成的物理原理,也没有考虑解决噪声放大的问题.针对上述问题,本文通过对图像降质的本质原因进行分析,提出一种基于渐进式双网络模型的低曝光图像增强方法,该方法包含图像增强模块以及图像去噪模块两个部分.对每个模块的构建也采用了渐进式的思想,考虑了图像由暗到亮的亮度变化,以及从粗到细的图像恢复过程,使增强后的结果更接近真实图像.为了更好地训练网络,本文构建了一种双向约束损失函数,从图像降质模型的正反两个方向使网络学习结果逼近真实数据,达到动态平衡.为了验证本文方法的有效性,本文与一些主流的方法从主观和客观两方面进行了实验对比,实验结果证明了本文方法得到的结果更接近真实图像,获得了更优的性能指标.When the traditional image enhancement methods enhanced the low-exposure image,they usually just considered the enhancement of brightness and ignored the problem of noise amplification.Besides,the current deep learning methods used the end to-end network to directly learn the mapping relationship between the low-exposure image and the normal image,ignoring the physical principle of the formation of the low-exposure image,and did not consider solving the problem of noise amplification.In order to solve these problems,this paper presents a low-exposure image enhancement method based on progressive dual network model by analyzing the essential causes of image degradation.The proposed method includes two parts:image enhancement module and image denoising module.The construction of each module also adopts the progressive idea by considering the image brightness change from dark to light and the image restoration from coarse to fine,so that the enhanced result is closer to the real image.Furthermore,to train the network better,a bidirectional constraint loss function is designed,which makes the learning result of network approach the real data from positive and negative directions of the image degradation model,and finally achieves dynamic balance.Experimental results show that the proposed method is more effective than some state-of-the-art enhancement methods from both subjective and objective evaluations.

关 键 词：低曝光图像增强 渐进式 双网络 双向约束损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]