一种车辆识别代号检测和识别的弱监督学习方法

A weakly supervised learning method for vehicle identification code detection and recognition

车辆识别代号对于车辆年检具有重要的意义。由于缺乏字符级标注,无法对车辆识别代号进行单字符风格校验。针对该问题,设计了一种单字符检测和识别框架,并对此框架提出了一种无须字符级标注的弱监督学习方法。首先,对VGG16-BN各个层次的特征信息进行融合,获得具有单字符位置信息与语义信息的融合特征图;其次,设计了一个字符检测分支和字符识别分支的网络结构,用于提取融合特征图中的单字符位置和语义信息;最后,利用文本长度和单字符类别信息,对所提框架在无字符级标注的车辆识别代号数据集上进行弱监督训练。实验结果表明,本文方法在车辆识别代号测试集上得到的检测Hmean数值达到0.964,单字符检测和识别准确率达到95.7%,具有很强的实用性。

The vehicle identification code(VIN)is of great significance to the annual vehicle inspection.However,due to the lack of character-level annotations,it is impossible to perform the single-character style check on the VIN.To solve this problem,a single-character detection and recognition framework for VIN is designed and a weakly supervised learning algorithm without character-level annotation is proposed for this framework.Firstly,the feature information of each level of VGG16-BN is fused to obtain a fusion feature map with single-character position information and semantic information.Secondly,a network structure for both the character detection branch and the character recognition branch is designed to extract the position and semantic information of a single character in the fusion feature map.Finally,using the text length and single-character category information,the proposed framework is weakly supervised on the vehicle identification code data set without character-level annotations.On the VIN test set,experimental results show that the proposed method realizes the Hmean score of 0.964 and a single-character detection and recognition accuracy rate of 95.7%,showing high practicability.