一种改进的车牌识别算法  

An improved license plate recognition algorithm

作　　者：赵伟[1] 胡文艺[1] 罗荔豪 黄元 ZHAO Wei;HU Wenyi;LUO Lihao;HUANG Yuan(School of Computer and Network Security,Chengdu University of Technology,Chengdu 610000,China)

机构地区：[1]成都理工大学计算机与网络安全学院,四川成都610000

出　　处：《电子设计工程》2023年第3期38-42,48,共6页Electronic Design Engineering

基　　金：教育部高等教育司协同育人项目(201802118013);地质灾害防治与地质环境保护国家重点实验室人才培养课题(SKLGP2018Z004)。

摘　　要：针对现如今车牌识别工作量大且识别速度和识别准确率低的问题,提出一种基于YOLOv4的识别算法,融合MobileNet网络结构修改其特征提取的主干网络,并加入黑盒预测模型。融合后的算法不仅大大缩小了自身的模型大小,而且相比原来的YOLOv4算法有着更高的识别准确率,同时省去了人工编码的工作。以本地实拍车牌作为数据集来训练和测试算法模型,以精确率Precision、召回率Recall、F1分数和平均准确率mAP作为评价指标依次对融合后的新模型以及YOLOv4进行评判。实验结果表明,MobileNetV1-YOLOv4实验效果相比YOLOv4更佳,mAP高达98.08%,而模型大小仅为49 MB,准确率提升了1.75%,模型缩小了80.48%,性能有着显著的提升。In view of the large workload of license plate recognition and the low recognition speed and recognition accuracy,proposes a recognition algorithm based on YOLOv4,which integrates the MobileNet network structure and modifies its feature extraction backbone network,and adds a black box prediction model. The fused algorithm not only greatly reduces the size of its own model,but also has a higher recognition accuracy rate than the original YOLOv4 algorithm,while eliminating the need for manual coding. The local real-photographed license plate is used as a data set to train and test the algorithm model,and the accuracy rate Precision,recall rate Recall,F1 score and average accuracy rate mAP are used as evaluation indicators to evaluate the fused new model and YOLOv4 in turn. The experimental results show that the experimental results of MobileNetV1-YOLOv4 are better than YOLOv4,mAP is as high as 98.08%,while the model is only 49 MB,the accuracy rate is increased by 1.75%,the model is reduced by 80.48%,and the performance is significantly improved.

关 键 词：图像识别 深度可分离卷积 MobileNet YOLOv4 

分 类 号：TN919.82[电子电信—通信与信息系统]