基于Transformer的车辆年款细粒度识别研究

Research on Fine-Grained Recognition of Vehicle Model Year Based on Transformer

视频监控场景下车辆年款信息抽取对城市数智化治理有着重要意义。为实现细粒度车辆年款的精准识别,首先,构建了覆盖多元采集条件及常见车辆年款的百万级场景数据集;其次,提出了基于Transformer的车辆年款细粒度特征高效提取器;最后,结合任务特点设计了层次标签多任务联合学习方法,获得兼容全局与局部的高鲁棒性特征。实验结果表明,提出的方法在场景数据集上的Top-1准确率达到95.79%,相较基于CNN的单任务方法有大幅提升。

Vehicle model year information extraction in video surveillance scenes is of great significance for urban digital intelligent governance.In order to achieve accurate identification of fine-grained vehicle model year,firstly,a mega scene dataset covering multiple collection conditions and common vehicle model year is constructed;secondly,an efficient fine-grained feature extractor of vehicle model year based on Transformer is proposed;finally,a hierarchical label multi task joint learning method is designed based on task characteristics to obtain high robustness features compatible with global and local features.The experimental results show that the Top-1 accuracy of the proposed method on the scene dataset reaches 95.79%,which is significantly improved compared with the single task method based on CNNs.