摘要
文中提出了一种新型基于轻量卷积神经网络的车辆车牌定位识别方法.采用基于改进的YOLOv3算法对车牌进行定位,基于免分割的轻量卷积神经网络LPRNet识别车牌字符.在车牌定位方面,改进YOLOv3网络的特征提取网络以降低设备要求,同时加入密集连接网络,增加对浅层特征信息的重复利用;在损失函数方面,引入DIOU损失函数,加快网络收敛以提高YOLOv3网络的定位精度;在车牌字符识别方面,采用基于免分割轻量卷积神经网络识别车牌字符,准确率高且保证了网络的轻量.结果表明:改进的YOLOv3网络算法的平均正确率降低了2.2%在CPU上检测速度达到了35帧/s,结合字符识别网络,总体检测速度达到27帧/s,满足实时性检测要求.
A new vehicle license plate location and recognition method based on lightweight convolutional neural network was proposed.The license plate is located based on the improved YOLOv3 algorithm,and the license plate characters were recognized based on the segmentation-free lightweight convolutional neural network LPRNet.In the aspect of license plate location,the feature extraction network of the improved YOLOv3 network was improved to reduce the equipment requirements,and the dense connection network was added to increase the reuse of shallow feature information.In the aspect of loss function,DIOU loss function was introduced to speed up the network convergence and improve the positioning accuracy of YOLOv3 network.In the aspect of license plate character recognition,the recognition of license plate characters based on segmentation-free lightweight convolutional neural network was adopted,which had high accuracy and ensures the lightweight of the network.The results show that the average accuracy of the improved YOLOv3 network algorithm is reduced by 2.2%,but the detection speed on the CPU reaches 35 fps.Combined with the character recognition network,the overall detection speed reaches 27 fps,which meets the requirements of real-time detection.
作者
程闯
梅磊
谭昕
CHENG Chuang;MEI Lei;TAN Xin(School of Intelligent Manufacturing,Jianghan University,Wuhan 430056,China)
出处
《武汉理工大学学报(交通科学与工程版)》
2023年第3期414-420,共7页
Journal of Wuhan University of Technology(Transportation Science & Engineering)
关键词
轻量化卷积神经网络
车牌识别与定位
密集连接
字符识别
lightweight convolutional neural network
recognition and location of license plate
dense connection
characters recognition