基于通道注意力与光照权重的无监督低照度图像增强  

Unsupervised Low Illumination Image Enhancement Based on Channel Attention and Illumination Weights

作　　者：杨猛 杜晓刚 张学军[3] 孙浩轩 YANG Meng;DU Xiaogang;ZHANG Xuejun;SUN Haoxuan(Shaanxi Joint Laboratory of Artificial Intelligence,Shaanxi University of Science&Technology;School of Electronic Informa-tion and Artificial Intelligence,Shaanxi University of Science&Technology,Xi'an 710021,China;School of Electronic and In-formation Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]陕西科技大学陕西省人工智能联合实验室 [2]陕西科技大学电子信息与人工智能学院,陕西西安710021 [3]兰州交通大学电子与信息工程学院,甘肃兰州730070

出　　处：《软件导刊》2024年第7期167-173,共7页Software Guide

基　　金：国家自然科学基金项目(61861024,62271296);甘肃省自然科学基金项目(21JR7RA282);陕西省重点研发计划项目(2021ZDLGY08-07)。

摘　　要：现有部分无监督低照度图像增强方法在增强图像曝光不足的区域时会降低其高光区域亮度,导致增强后的图像出现伪影;单一的TV损失既无法区别照明特征图的细节,还会忽略照明特征图边缘明暗度差异突出的地区,导致光晕现象的产生。为此,提出一种基于通道注意力与光照权重的无监督低照度图像增强方法 VARRNet。首先,VARRNet将图像转化为HSV空间,将V空间与Retinex理论结合以避免损失信息;其次,为了防止在亮度增强过程中生成伪影,设计了一个亮度估计网络引入通道注意力ECA分配输入特征图的权重,以恢复曝光不足区域的亮度,并有效保持高光区域的亮度;最后,在亮度估计网络中结合TV损失与光照分量权重来保留增强后特征图的丰富细节信息,消除强边缘处的光晕。在与当前流行的5个低照度图像增强方法进行比较实验发现,VARRNet在亮度增强、细节保留、色彩恢复、伪影抑制和光晕去除等方面均取得了更好的可视化效果。Some existing unsupervised low light image enhancement methods may reduce the brightness of highlights in areas with insufficient image exposure,resulting in artifacts in the enhanced image;A single TV loss cannot distinguish the details of the lighting feature map,and it will also ignore areas with prominent differences in brightness at the edges of the lighting feature map,leading to the occurrence of halo phenomena.To this end,a unsupervised low light image enhancement method VARRNet based on channel attention and lighting weight is proposed.Firstly,VARRNet converts images into HSV space and combines V space with Retinex theory to avoid information loss;Secondly,in order to prevent the generation of artifacts during the brightness enhancement process,a brightness estimation network was designed to introduce channel attention ECA to allocate the weights of input feature maps,in order to restore the brightness of underexposed areas and effectively maintain the brightness of highlight areas;Finally,in the brightness estimation network,TV loss and lighting component weight are combined to preserve the rich detail information of the enhanced feature map and eliminate halos at strong edges.Compared with five popular low light image enhancement methods,VARRNet achieved better visualization results in brightness enhancement,detail preservation,color restoration,artifact suppression,and halo removal.

关 键 词：无监督学习 RETINEX 低照度图像增强 通道注意力 照明平滑度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]