MIRNet-Plus:基于丰富特征学习的低光图像增强改进方法  

MIRNet-Plus:a Low Light Image Enhancement Improved Method Based on Rich Feature Learning

作　　者：罗林 余联想 郑明魁[1,2] LUO Lin;YU Lianxiang;ZHENG Mingkui(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362200,China;College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学先进制造学院,福建泉州362200 [2]福州大学物理与信息工程学院,福州350108

出　　处：《小型微型计算机系统》2024年第3期664-669,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61902071)资助;福建省自然科学基金项目(2020J01466)资助;闽都创新实验室主任基金项目(2021ZR151)资助.

摘　　要：图像增强是一种基础的计算机视觉任务,从低光图像中恢复出高质量的明亮图像是业界正在攻克的问题.近年来,以卷积神经网络(CNN)为主导的图像恢复技术取得了重大进展.对于低光图像增强,本方法使用双重选择核融合(Double SKFF)方法,通过增强中间层不同分辨率信息的交流能力以获得更多上下文信息以及空间信息;同时设计了一个深度注意模块(Depthwise Attention Module,DWM),用来共享张量中的特征信息,对原有特征进行补充,获取更加丰富的特征信息.同时本方法还引入了多颜色空间神经修饰块,用来在3种不同的颜色空间(Lab,RGB,HSV)中联合训练,以期望获得更好的图像增强结果.本文提出的MIRNet-Plus在原有的基础方法上PSNR获得了5.3%的提高,由23.73dB提升到24.98dB.Image enhancement is a fundamental computer vision task,and recovering high-quality bright images from low-light images is a problem being tackled by the industry.In recent years,significant progress has been made in image restoration techniques led by convolutional neural networks(CNN).For low-light image enhancement,this method uses double selective kernel fusion(Double SKFF)to obtain more contextual information and spatial information by enhancing the communication ability of different resolution information in the middle layer;and a Depthwise Attention Module(DWM)is designed to share the feature information in the tensor to obtain more contextual information,also,this module is designed to share the feature information in the tensor to supplement the original features and obtain richer feature information.A multi-color spatial neural modification block is also introduced to jointly train in three different color spaces(Lab,RGB,HSV)in order to obtain better image enhancement results.In this paper,the proposed MIRNet-Plus obtains a 5.3%improvement in PSNR from 23.73 dB to 24.98 dB over the original base method.

关 键 词：图像增强 低光图像 卷积神经网络 深度注意模块 多颜色空间神经修饰块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]