基于迁移学习的发票号码识别研究  

Research on Invoice Number Recognition Based on Transfer Learning

作　　者：黄为新 陶杨 张继超 苏笛 牛砚 HUANG Wei-xin;TAO Yang;ZHANG Ji-chao;SU Di;NIU Yan(College of Computer Science and Technology,Jilin University;College of Math,Jilin University,Changchun 130012,China)

机构地区：[1]吉林大学计算机科学与技术学院 [2]吉林大学数学学院,吉林长春130012

出　　处：《软件导刊》2021年第6期45-48,共4页Software Guide

基　　金：吉林大学校级大学生创新创业训练项目(201910183X365)。

摘　　要：财务系统中的发票管理长期以来严重依靠人力,并存在一定的失误率,且目前一些发票识别方法存在准确率不高、数字识别实时性不足等问题。鉴于此,基于已有较成熟的手写体数字识别工作,从迁移学习思想和方法入手,设计一种能准确快速识别发票号码的新方法。选取迁移学习的3种具体方法,即Tradaboost、微调和卷积神经网络提取特征并运用于SVM,分别应用于数字识别模块。通过大量测试和比对,结果表明,第3种方法即采用CNN结合SVM的算法在实验中能达到99.75%的准确率,且识别速度快,具有较好的稳定性、鲁棒性。Invoice management in the financial system has long been heavily dependent on manpower,which is time-consuming and laborious,and has a certain error rate.This area involves the financial industry,which is very sensitive to digital information,and requires low misinformation rate and high reliability.At present,there are mainly two methods for the problem of invoice recognition,they are recognizing the invoice number through the structural characteristics of the number and using the convolutional neural network to recognize the number.However,these methods are not accurate enough and the real-time performance of the number recognition is not high enough.Because the work for handwritten number recognition has been mature,this paper starts with the idea and method of transfer learning and applies it to the work of invoice print number recognition,and designs a new method of accurately fast identification of invoice number,so that the human from tedious digital finishing work,reduce the manpower investment.Three specific transfer learning methods are selected,including Tradaboost,fine-tuning and using convolution neural network to extract features applied to SVM which are applied into the digital recognition module respectively.By a lot of testing,the results show that the third method of CNN combined with SVM can reach 99.75% accuracy in the experiments and the recognition speed is fast,which proves the algorithm has better stability and robustness.

关 键 词：发票识别 印刷数字识别 字符分割 迁移学习 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]