基于深度学习的非约束场景多车牌识别方法研究  

Research on Unconstrained Scene Multiple License Plate Recognition Method Based on Deep Learning

作　　者：张振威 王红成[1] ZHANG Zhenwei;WANG Hongcheng(School of Electrical Engineering and Intelligentization,Dongguan University of Technology,Dongguan 523808,China;School of Computer Science and Technology,Dongguan University of Technology,Dongguan 523808,China)

机构地区：[1]东莞理工学院电信工程与智能化学院,广东东莞523808 [2]东莞理工学院计算机科学与技术学院,广东东莞523808

出　　处：《东莞理工学院学报》2024年第1期20-27,共8页Journal of Dongguan University of Technology

基　　金：东莞市科技特派员项目(20201800500232);东莞理工学院2020年科技产业创新服务专项行动成果育成小分队项目。

摘　　要：为解决非约束场景下的车牌识别精度问题,提出一种能够适应非约束场景并识别不同类型单车牌以及多车牌的自动车牌识别模型。该模型应用数据迁移技术,通过YOLOv5检测车辆并通过后处理筛选有效车辆目标,经检测并矫正后,通过ResNet18以及双向长短期记忆网络BLSTM网络结合连接时序分类损失CTC识别车牌字符。在模型训练过程中使用数据增强技术进一步提高了模型性能。该模型在CCPD以及AOLP的多个子数据集上进行了测试,展现出了优于其它方法识别精度和识别速度。To solve the problem of license plate recognition accuracy in unconstrained scenarios,an automatic license plate recognition model that can adapt to unconstrained scenarios and recognize different types of single license plates and multiple license plates is proposed.The model uses data migration technology to detect vehicles through YOLOv5 and screen effective vehicle targets through post-processing.After detection and correction,the license plate characters are recognized through ResNet18 and BLSTM networks combined with CTC loss.Using data augmentation techniques during model training further improves model performance.After testing on multiple sub-datasets of CCPD and AOLP,the presented model shows better recognition accuracy and recognition speed than other methods.

关 键 词：车牌识别 非约束场景 深度学习 数据增强 多车牌 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]