Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (18): 105-118.DOI: 10.3778/j.issn.1002-8331.2211-0245

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Multi-Scale Deformable Transformer for Banknote Serial Number Recognition

ZHANG Kaisheng, LI Xuyang   

  1. School of Electrical and Control Engineering, Shaanxi University of Science & Technology, Xi’an 710021, China
  • Online:2023-09-15 Published:2023-09-15

多尺度可变形Transformer纸币序列号识别

张开生,李旭洋   

  1. 陕西科技大学 电气与控制工程学院,西安 710021

Abstract: The serial number recognition system of banknotes plays an important role in the supervision of the circulation of banknotes, and the banknotes will be deformed or contaminated in the process of actual use, resulting in the irregular text of the serial number of the banknotes. In order to meet the identification requirements of irregular banknote serial numbers, a multi-scale deformable Transformer-based identification method for banknote serial numbers is proposed. Through the self-built banknote serial number detection platform, the image of the banknote serial number is obtained and transmitted to the computer. The feature map of the serial number image is extracted through the backbone network and transmitted to the encoder. The multi-scale feature information of the text is further extracted through the multi-scale deformable attention mechanism, and then the thick bounding box of the text is extracted by the candidate box generator using the polygon bounding box detection mechanism. The regression training of the polygon bounding box coordinates in the position decoder is guided. The character decoder performs character prediction while the position decoder predicts the text bounding box, and finally the text recognition result of the serial number of the banknote is output. The experimental results show that this method can meet the needs of online detection and identification of banknote serial numbers.

Key words: banknote serial number, deep learning, Transformer, character recognition

摘要: 纸币序列号识别系统在纸币流通的监管中扮演着重要角色,而纸币在实际使用的过程中会产生变形或受到污染导致纸币序列号呈现出不规则文本的特点。为满足不规则纸币序列号的识别需求,提出基于多尺度可变形Transformer的纸币序列号的识别方法。通过自主搭建的纸币序列号检测平台获取纸币序列号图像传输至计算机。序列号图像通过骨干网络提取特征图并传输至编码器,通过多尺度可变形注意力机制进一步提取文本的多尺度特征信息,随后采用多边形边界框检测机制,经候选框生成器提取文本的粗边界框,引导位置解码器中的多边形边界框坐标的回归训练,字符解码器在位置解码器预测文本边界框的同时进行字符预测,最终输出纸币序列号文本识别结果。实验结果表明,该方法能够满足纸币序列号在线检测与识别的需求。

关键词: 纸币序列号, 深度学习, Transformer, 字符识别