Research On Deep Learning-Based Banknote Serial Number Recognition

Deep learning has been widely used in image classification,target detection,image segmentation,and speech recognition in recent years and has achieved remarkable results.Convolutional neural network is an important network in the field of deep learning,and has many outstanding features that cannot be achieved by traditional classification algorithms.The character recognition problem has also been widely studied in recent years,and its scope is wider,including license plate recognition,handwriting recognition,identity card bank card character recognition,banknote serial number identification,and the like.In this article,the deep learning is applied to character recognition of serial number of banknotes according to the characteristics of serial number character data of RMB banknotes in circulation.the data sample used in this article is the fifth edition of the RMB banknote serial number image,and then it is subjected to character preprocessing such as character segmentation,scale conversion,sample amplification,and label tagging.A template matching segmentation algorithm based on prior knowledge is proposed,which has better stability for image segmentation with poor background sequence number.This article constructed a deep convolutional neural network CNN1 for banknote serial number character recognition,and it achieved good results.At the same time,for the problem of misjudgment,a convolutional neural network structure LGCNN1 combining global and local features is proposed,which effectively improves the accuracy of network recognition.On this basis,dilated convolution was introduced,and a deep dilated convolutional neural network DCNN1 for the serial number character recognition of banknotes studied in this article was constructed.Compared with the ordinary convolutional neural network,the time-consuming difference was not much more but a better recognition result was achieved.And the inflated convolutional neural network structure LGDCNN1 combining global and local features was constructed,which made the recognition rate of the characters 99.995%.