基于改进UNet网络的车牌分割算法研究  

Research on license plate segmentation algorithm based on improved UNet network

作　　者：杨创富 张昕 海燕 周飞 陈荣演 邱天 YANG Chuangfu;ZHANG Xin;HAI Yan;ZHOU Fei;CHEN Rongyan;QIU Tian(Department of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China;Jiangmen Fire Rescue Detachment,Jiangmen Fire Bureau,Jiangmen 529020,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020 [2]江门市消防局江门市消防救援支队,广东江门529020

出　　处：《电子设计工程》2023年第21期186-190,195,共6页Electronic Design Engineering

基　　金：广东省重点领域研发计划(2020B0101030002);2021年江门市创新实践博士后课题研究资助项目(JMBSH2021B04)。

摘　　要：针对传统图像算法在定位分割车牌时易受复杂环境因素影响,为提升分割准确率,采用深度学习UNet网络模型进行改进研究,对其增加注意力cSE模块,以增强网络对特征的提炼能力,与其他主流分割网络进行对比实验。实验结果表明,在自建车牌数据集上经过100轮训练后,该文改进方法中的各种指标系数表现最好,其中在测试集上的Dice、Miou指标分别比经典UNet网络提高了16%、15%,较FCN和CeNet等分割网络有较大幅度的提升,较主流的UNet改进网络也有不同程度的提升,证明了cSE-UNet网络能够提高车牌图像分割的准确率,是一种有效的改进网络模型。In view of the fact that traditional image algorithms are easily affected by complex environmental factors when locating and segmenting license plates,in order to improve the segmentation accuracy,adopts the deep learning UNet network model for improvement and research,and adds the Attention cSE module to it,so as to enhance the ability of the network to extract features,and then compares it with other mainstream segmentation networks for experimental analysis.The experimental results show that after 100 rounds of training on the self built license plate dataset,the performance of various index coefficients in the improved method is the best.The Dice and Miou indexes on the testingdataset are 16% and 15% higher than those of the classical UNet network respectively.Compared with segmentation networks such as FCN and CeNet,the UNet network has also been improved to varying degrees,which proves that the cSE-UNet network can improve the accuracy of license plate image segmentation and is an effective improved network model.

关 键 词：车牌分割 注意力机制 UNet网络改进 cSE模块 

分 类 号：TN99[电子电信—信号与信息处理]