面向数据脱敏的交通场景车牌检测方法  

A Detection Method of Traffic Scene License Plate for Data Desensitization

作　　者：应申[1] 曾卓源 张纪元 YING Shen;ZENG Zhuoyuan;ZHANG Jiyuan(School of Resource and Environmental Sciences,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学资源与环境科学学院,武汉430079

出　　处：《交通信息与安全》2024年第6期84-94,共11页Journal of Transport Information and Safety

基　　金：国家重点研发计划项目(2021YFB2501101);湖北重大科技攻关项目(2023BAA017)资助。

摘　　要：从车载相机图像中快速准确地检测车牌对于保护交通敏感信息具有重要意义。针对传统YOLOv8算法对交通场景下车牌检测存在的小目标特征提取能力弱、背景信息误检等问题,研究了基于改进YOLOv8的交通场景车牌检测方法TLP-YOLO。为增强主干网络提取图像特征信息能力,引入多尺度注意力模块(efficient multi-scale attention,EMA),加强对不同尺度目标区域的关注,提升模型对背景信息的甄别能力;提出跳阶加权特征金字塔网络,丰富小目标特征,通过跳阶连接和加权融合方式,避免特征金字塔结构不同层级间的信息损失,提升模型的多尺度特征融合能力;为降低模型计算量并保持检测精度,对检测头进行轻量化处理,引入部分卷积(partial convolution,PConv)和逐点卷积(pointwise,PWConv)模块代替常规卷积结构,以高效利用空间特征,避免冗余计算。基于中国城市停车数据集(Chinese city parking dataset,CCPD)和中国道路车牌数据集(Chinese road plate dataset,CRPD)构建用于验证模型性能的多交通场景数据集并进行了算法验证。实验结果表明:(1)TLP-YOLO的AP50-95和AP70分别达到了83.6%,和97.7%,相较于基准算法YOLOv8,其AP50-95和AP70分别提高了2%和0.8%。(2)TLP-YOLO模型的计算复杂度(floating point operations,FLOPs)为7.5 G,模型参数量为1.67 M,检测速度达到了101 fps,相较于基准算法YOLOv8,计算复杂度降低了8.5%,模型参数量减少了45%,平均检测速度则与之相当。改进后的算法能够在保证模型轻量化的同时,提高检测精度,满足车载设备对交通场景中车牌检测准确性与部署要求。Rapid and accurate detection of license plates in vehicle-based images is significant for protecting priva-cy information in smart transportation.However,the original YOLOv8 algorithm has limitations on the license plate detection in traffic scenes,such as weak feature extraction ability of small targets and misdetection of background information,etc.To fill these gaps,an improved traffic scene license plate detection method based on YOLOv8(TLP-YOLO)is proposed.The efficient multi-scale attention(EMA)module is adopted to enhance the ability of the backbone network to extract image characteristics.It makes the backbone network pay more attention to target re-gions of different scales and improves the recognition ability of the model to background information.A new feature pyramid network with skip connection and weighted fusion(SW-FPN)is designed.It enriches the features of small targets and avoid the information loss between different levels of the feature pyramid network,which improving the multi-scale feature fusion ability.In order to reduce the floating-point operations(FLOPs)and maintain the detec-tion accuracy,the partial convolution(PConv)and pointwise convolution(PWConv)modules are introduced to re-place the conventional convolution structure in detection head,which reduces redundant calculations and improves the utilization efficiency of spatial features.Based on Chinese city parking dataset(CCPD)and Chinese road plate dataset(CRPD),a dataset with multiple traffic scenes is constructed to verify the property of the model.Experimen-tal results show that:①The average precision(IOU changes from 0.5 to 0.95)of the proposed network is 83.6%,which is 2%higher than that of YOLOv8.The average precision(IOU is 0.7)of the proposed network is 97.7%,which is 0.8%higher than that of YOLOv8.②The FLOPs of TLP-YOLO model is 7.5 G,the number of parameters is 1.67 M,and the detection speed reaches 101 fps.In comparison to the original YOLOv8,the FLOPs and the num-ber of parameters is reduced by 8%and 45%,the detection

关 键 词：数据脱敏 车牌检测 YOLOv8 交通场景 特征融合 高效多尺度注意力 

分 类 号：U412[交通运输工程—道路与铁道工程]