Design And Implementation Of An Automatic Invoice Recognition System Based On Deep Learning

In recent years,with the rapid development of China’s economy,various commodity transactions are becoming more and more frequent.Invoices,as the voucher for transaction activities and the basic basis for accounting by the financial department,play a significant role in our country’s economic life.In the traditional economic life,the management process of invoices is also cumbersome and complicated,which consumes a lot of manpower and material resources,and leads to low efficiency.With the development of science and technology,especially the development of related technologies in the field of deep learning,it has provided the possibility to further improve the efficiency of invoice management,and various invoice management systems have emerged.This thesis combines related technologies in traditional fields with that in deep learning to design and implement an automatic invoice recognition system,which can further improve the efficiency of invoice management.This thesis mainly introduces the overall framework of this automatic invoice recognition system and the functions and implementation methods of each module,especially the invoice classification module,text detection module and text recognition module.Among them,the classification method based on template matching and the classification method based on convolutional neural network are used in the invoice classification module,and the two are compared and analyzed.In addition,this thesis designs a multi-region localization method for the text detection module and the text recognition module,which can better combine the two modules,and at the same time,it can reduce the requirement of the detection accuracy of the text detection algorithm,and solve the invoice text printing offset and so on.Finally,this thesis focuses on the implementation of the text detection module and the text recognition module.The CTPN algorithm and the CRNN algorithm are used respectively.And this thesis improved the CRNN algorithm,so that it can identify the space key,and better apply to the multi-region localization method.Finally,we built an interface for the entire system,and tested the system,and analyzed the pros and cons of the automatic invoice recognition system designed and implemented in this thesis,and put forward the direction for improvement.