Research On Invoice Classification Based On Deep Learning And Implementation Of Intelligent Reimbursement System

With the rapid growth of China's economy,the demand for invoices from all walks of life has also increased.Invoices are an important voucher for the financial department of each work unit to check economic activities and reimburse.The current invoice reimbursement process is cumbersome,inefficient,and requires a large number of financial personnel,which greatly wastes social resources.Therefore,the intelligentization of invoice reimbursement is the general trend.This paper uses computer vision software library and deep learning methods to complete the tasks of correction and classification of invoice images,and combines hardware equipment to realize the entire process of invoice reimbursement.First of all,this article completes the correction of different oblique images according to the different characteristics of the invoice.For train tickets with obvious contours,this article firstly performs grayscale processing and threshold processing on the image,then obtains the contour of the invoice image through the computer vision software library,and extracts the contour containing the image,and finally corrects the image according to the contour.For value-added tax invoices with obvious straight lines in the invoice,this article uses Hough transform to detect the long straight lines in the invoice image,and count the inclination angles of the detected lines to find the average value of the inclination angle,and finally the image Correction.Secondly,this paper proposes an invoice classification model based on the residual network,which borrows the special structure of the residual network in structure.In the process of model training,this article uses a variety of supervised data enhancement methods such as flipping and random erasure,as well as a variety of unsupervised data enhancement methods such as image style conversion,which effectively improves the generalization of the model,so that the value added tax can be reduced.Invoices,train tickets and taxi invoices are accurately classified.Through experimental comparison,this model has the best classification accuracy rate,and its classification accuracy rate reaches 98.81%,and the model can not only be applied to invoice classification,but also can modify the training process of the model so that it can be applied to other image classifications.On mission.Finally,this article combines hardware equipment to realize the whole process from reimbursement confirmation to invoice scanning,identification,classification,and printing of reimbursement vouchers.In the process of obtaining ID card information,through the second generation ID card reader During the second development,the key information such as ID card photos,names and other key information needed by the system was efficiently obtained.In the stage of printing the reimbursement voucher,the reimbursement information and all the various types of tickets scanned during the reimbursement process are summarized,and the reimbursement time is recorded,and the above information is printed out through a thermal printer to complete the entire reimbursement process.