融合双层注意力网络的端到端老挝车牌照识别方法

An end-to-end Laos ehicle license plate recognition method integrating double layer attention networks

在中老道路互通大背景下,老挝车牌照识别研究对中国跨境车辆管理十分重要,但现有的单行车牌照识别方法无法直接应用于老挝双行车牌照识别任务中.针对老挝车牌照上行省份字符排列紧密、难以分割和下行辅音字符相似度高、难以识别的问题,结合分割的思想提出一种融合双层注意力网络的端到端老挝车牌照识别方法.通过通道及空间注意力提取并加强上行省份特征和下行字符特征表示;将分类思想应用于省份信息获取,有效地处理因字符粘连而无法做单字符识别的问题;使用序列标注的方法缓解相似字符识别困难,提高字符识别准确率.实验结果表明,提出方法相比基线模型,准确率提升了0.8个百分点,达到92.7%.

Against the backdrop of road connectivity between China and Laos,research on license plate recognition in Laos is crucial for cross-border vehicle management in China.However,existing single license plate recognition methods cannot be directly applied to dual license plate recognition tasks in Laos.This paper proposes an end-to-end Lao car license plate recognition method that integrates a two-layer attention network,in response to the problems of tight arrangement of characters in the upstream provinces of the Lao car license plate,difficulty in segmentation,and high similarity and difficulty in recognizing the downstream consonant characters.Extracting and enhancing the representation of upstream province features and downstream character features through channel and spatial attention.Applying the idea of classification to provincial information acquisition,effectively addressing the problem of single character recognition not being possible due to character adhesion.Using sequence annotation to alleviate the difficulty of similar character recognition and improve the accuracy of character recognition.The experimental results show that the proposed method improves the accuracy of the baseline model by 0.8 percentage points,reaching 92.7%.