第39卷第2期2024年2月Vol.39No.2Feb.2024液晶与显示ChineseJournalofLiquidCrystalsandDisplays基于深度学习的雾霾天气下的车牌号码识别方法杨云*,王静,姜佳乐(陕西科技大学电子信息与人工智能学院,陕西西安710021)摘要:针对雾霾天气下车牌识别存在的精确度低、漏检等问题,提出了一种基于深度学习的雾霾天气下的车牌号码识别方法。首先用AOD-Net算法对车辆图像进行去雾预处理。然后,基于YOLOv5网络设计一种车牌检测网络ACG_YOLOv5s。ACG_YOLOv5s是在YOLOv5s网络的基础上,融入CBAM注意力机制,提高模型的抗干扰能力;引入自适应特征融合网络ASFF,根据模型自适应学习到的权重赋予网络不同特征层不同比重的权值,从而突出重要特征信息;使用Ghost卷积模块替换传统卷积,在保证模型效果的同时减少了网络训练过程中的参数量。最后通过LPRNet对检测到的车牌图像进行识别。实验结果表明,改进后的ACG_YOLOv5s网络车牌检测准确率达到99.6%,LPRNet识别准确率达96%且内存占比小。实验证明AOD-Net算法和YOLO算法结合可更加有效地检测雾霾天气下车牌图像中的车牌号码。关键词:车牌号码识别;AOD-Net算法;YOLOv5网络;注意力机制中图分类号:TP391文献标识码:Adoi:10.37188/CJLCD.2023-0123MethodofvehiclelicenseplaterecognitioninhazeweatherbasedondeeplearningYANGYun*,WANGJing,JIANGJiale(SchoolofElectronicInformationandArtificialIntelligence,ShaanxiUniversityofScienceandTechnology,Xi'an710021,China)Abstract:Adeeplearningbasedlicenseplatenumberrecognitionmethodisproposedtoaddresstheissuesoflowaccuracyandmisseddetectioninlicenseplaterecognitionunderhazeweather.Firstly,theAOD-Netalgorithmisusedtopre-processthevehicleimagefordefogging.Then,alicenseplatedetectionnetworkACG_YOLOv5sisdesignedbasedonYOLOv5network.ACG_YOLOv5sintegratesCBAMattentionmechanismonthebasisofYOLOv5snetworktoimprovethemodel’santi-interferenceability.Anadaptivefeaturefusionnetwork(ASFF)isintroduced,whichassignsweightstodifferentfeaturelayersofthenetworkbasedontheweightsadaptivelylearnedbythemodel,therebyhighlightingimportantfeatureinformation.ThetraditionalconvolutionisreplacedwithGhostconvolutionmoduleandthenumberofparametersduringnetworktrainingisreducedwhileensuringmodel...