Banknotes Serial Number Coding Recognition
Xu Ruru, Jungang An, Su Liandeng, Xinli Min · 2019
Banknotes serial number coding is the main basis of bank's supervision, currency circulation and authenticity verification, the traditional manual identification method can hardly meet the current needs of currency circulation and security control. This paper solves the problem of banknotes serial number coding recognition based on deep learning algorithm. Our algorithm is composed of text localization and text recognition. The aim of the text localization phase is to get the region of the serial number coding. Based on YOLOv3 model, this paper replaces original convolution module with depthwise separable convolution. The idea of Inception is adopted in subsample stage. K-means algorithm is used to re-cluster width and height of serial number coding region to adjust anchor value, thus, improving the localization accuracy. The aim of the text recognition phase is to get the serial number coding of the banknotes, CNN+LSTM+CTC is taken as the model for text recognition. In this paper, the fourth edition of RMB data set is used for serial number coding recognition, 99.38% accuracy is obtained in experimental results.