基于深度学习算法的车牌检测系统设计  

Design of License Plate Detection System Based on Deep Learning Technology

作　　者：王东升 聂建军[1] WANG Dongsheng;NIE Jianjun(School of Mechatronics Engineering,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院机电学院,河南郑州450007

出　　处：《成组技术与生产现代化》2024年第3期27-35,共9页Group Technology & Production Modernization

摘　　要：为实现汽车车牌的自动读取,设计了一种基于深度学习的轻量化车牌检测系统。在YOLOv8网络模型的基础上,用MobileNetV3网络更换主干网络,减少了模型的参数量,提升了车牌检测速度;引入全维度动态卷积来调整Neck模块的扩展率,提升了车牌检测精度。在用目标检测网络对车牌识别后,利用PaddleOCR软件进行了字符识别。利用PyQt5软件对检测系统的操作界面进行设计,并对软件的操作过程进行了说明。在选用的中国城市停车数据集(Chinese City Parking Dataset,CCPD)上进行了实验验证。验证结果表明:改进后网络模型的运算复杂度GFLOPs为7.8,检测平均精度mAP50为89.77%,运算速度FPS为86.1帧/s。相较于现有车牌检测算法所用网络模型,改进后网络模型有效地兼顾了轻量化和检测精度要求,可以满足汽车车牌实时检测的需要。Aiming at the purpose of automatic reading of vehicle license plates,this paper proposes a lightweight detection method based on an improved YOLOv8.Firstly,the backbone network is replaced with MobileNetV3,significantly reducing the number of model parameters and improving the detection speed of the model.Secondly,by introducing full-dimensional dynamic convolution to replace the regular Conv in the Neck and adjusting the neck expansion rate,the detection accuracy is improved while reducing the number of model parameters.Afterward,PaddleOCR is used to recognize the characters of the vehicle license plates identified by the target detection network.Finally,this paper designs the operation interface of the detection system using PyQt5 and demonstrates the operation of the software.Experimental verification was performed on the selected CCPD dataset,and the results showed that the computational complexity GFLOPs of the improved model was 7.8,with an mAP 50 of 89.77%on the test set and an FPS of 86.1 frames/s.Compared to existing license plate detection algorithms,this model effectively balances lightweight and detection accuracy,meeting the needs of real-time vehicle license plate detection.

关 键 词：深度学习 车牌检测 网络模型 YOLOv8 MobileNetV3 系统设计 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]