基于注意力引导数据增强的车型识别  

Vehicle Type Recognition Based on Attention Guided Data Augmentation

作　　者：孙伟[1,2] 常鹏帅 戴亮[1] 张小瑞[2,3,4] 陈旋 代广昭 SUN Wei;CHANG Pengshuai;DAI Liang;ZHANG Xiaorui;CHEN Xuan;DAI Guangzhao(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China;Jiangsu Province Collaborative Innovation Center on Atmospheric Environment and Equipment Technology,Nanjing University of Information Science and Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics of Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学自动化学院,南京210044 [2]南京信息工程大学江苏省大气环境与装备技术协同创新中心,南京210044 [3]南京信息工程大学数字取证教育部工程研究中心,南京210044 [4]南京信息工程大学计算机与软件学院,南京210044

出　　处：《计算机工程》2022年第7期300-306,共7页Computer Engineering

基　　金：国家自然科学基金(61304205);江苏省自然科学基金(BK20191401,BK20201136);江苏省研究生科研与实践创新计划项目(SJCX21_0363);大学生创新创业训练项目(202010300290,202010300211,202010300116E)。

摘　　要：车型识别在智能交通系统中发挥着重要作用。受车辆数据不足、车辆类间差异小等因素的影响,传统车型识别方法未充分利用车辆鉴别性区域的特征,导致识别准确率降低。提出一种基于注意力模块引导数据增强的车型识别方法。将ResNet-50作为骨干网络提取车辆特征,同时在网络的每个残差块后均嵌入坐标注意力模块,编码成一对方向感知和位置敏感的注意力图,以增强车辆鉴别性区域的特征表达。在此基础上,利用双线性注意力汇集操作生成增强特征图,通过对增强特征图进行注意力裁剪和注意力擦除,获取具有强鉴别性的增强数据。在Stanford Cars车辆数据集上的实验结果验证了该方法的有效性,结果表明,该方法的车型识别准确率达到94.86%,与RA-CNN、MA-CNN、WS-DAN+Inception-v3等方法相比,能够有效提高车型识别准确率和数据增强效率。Vehicle type recognition plays an important role in intelligent transportation systems.Owing to the lack of vehicle data and small differences between vehicle classes,traditional vehicle type recognition do not make full use of the features of the vehicle discriminant area,resulting in a reduction in recognition accuracy. This study proposes a vehicle type recognition method based on attention guided data augmentation. In this method,ResNet-50 is used as the backbone network to extract vehicle features.Simultaneously,a Coordinate Attention(CA)module is embedded behind each residual block of the network to encode a pair of direction-aware and position-sensitive attention diagrams to enhance the feature representation of the vehicle discriminant area. On this basis,the bilinear attention-gathering operation is used to effectively obtain the enhanced feature image. Through attention cropping and erasure of the enhanced feature map,enhanced data with strong discrimination are obtained. The results on the Stanford Cars vehicle dataset verify the effectiveness of this method. The results showed that the accuracy of vehicle type recognition of this method reaches 94.86%. Compared with RA-CNN,MA-CNN,WS-DAN+Inception-v3,and other methods,it can effectively improve the accuracy of vehicle type recognition and efficiency of data augmentation.

关 键 词：车型识别 坐标注意力 数据增强 注意力裁剪 注意力擦除 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]