Research And Implementation Of Small-scale RMB Serial Number Recognition System

The RMB is the legal currency of China.The serial number on RMB banknotes is unique and plays an important role in the printing,storage and circulation of banknotes.In order to effectively solve the problems of proof and responsibility of banks involved in counterfeit banknote disputes,and improve the credibility of financial institutions,the People's Bank of China has proposed the requirement that financial instruments need to have serial number identification capability in the past two years.Based on the characteristics of RMB serial character,combined with embedded software,electronic hardware and image processing technology,this paper implements a set of small-scale currency counter system,which can quickly and stably identify the serial number on RMB.In terms of system architecture,this paper proposes improvements in software architecture and hardware architecture.The hardware architecture adopts multi-core ARM + FPGA + A/D + CIS architecture different from the traditional DSP + MCU + FPGA + A/D + CIS architecture,which not only improves the computing performance of the whole machine but also has higher cost performance;this paper uses the hardware FPGA program to achieve pre-processing of image brightness equalization compensation,which can save CPU resources.The software architecture is based on embedded Linux system,uses V4L2 to build the banknote image capture driver,builds applications with multi-threading,and builds a human-machine user interface with QT Graphical User Interface and adds RT-linux patch,which can improve the real-time performance of the system.In the process of image recognition algorithm of the serial number,this paper has been taking into account the needs of the system for banknotes counting speed and the limitation of system computing resources.Firstly this paper implements a fast algorithm for the positioning and segmentation of the serial number characters,and then compares and implements two kinds of serial number recognition methods.The first method is to use the SVM one versus rest multi-classifier to classify numbers after extracting the HOG features vector.In the first method,after SVM sample training and testing,the recognition rate is slightly lower,which does not meet the GB16999-2010 requirements.The second method of identification is to implement a convolutional neural network based on LeNet-5.The improved network structure includes two convolutional layers,one pooling layer,two fully connected layers,and uses Softmax as the output layer of the network.In the second method,after training a large number of samples,the recognition rate is obviously higher than the first method and meets the GB16999-2010 requirements.The system is put into mass production after debugging and testing in software,hardware systems and image algorithms.The system has excellent number recognition performance,economic cost advantage and cutting-edge technology advantages.