Research On Intelligent Financial Reimbursement System Method Based On Deep Learning

Recently,with the rapid development of the social economy,the invoices usage is increasing dramatically.As vouchers for various economic activities,invoices are essential parts of the financial reimbursement process.The conventional procedure of reimbursement is complicated,inefficient,and mainly manual,which wastes much human and material resources.This paper adppts conventional image processing and novel deep learning methods to complete the core tasks,such as classification,detection and identification of the invoice images.Firstly,this thesis proposes a template matching based method for identifying the scanned invoice images,which mainly includes four steps,such as image preprocessing,template matching,optical character recognition and information exporting.Specifically,the image preprocessing step performs secondary rotation and edge cutting operators on the original invoice image to obtain the main image without useless background,using edge detection and contour extraction methods.The template matching step compares the invoice image with a selection of template images prepared in advance to extract the text regions.The optical character recognition step obtains the character information from the extracted text regions.The information exporting step corrects and stores the recognized character information.Six methods of template matching are tested in the experiments,and it is found that the method using normalized correlation coeffcient matching achieves the best performance.The method achieves 95.02 % in accuracy,and it takes 8.6 milliseconds to identify twenty regions in one invoice image.Secondly,this thesis proposes a method based on convolutional neural network to classify the invoice images taken by smart phones.The proposed model is improved from VGG-16 model,and it is simplified according to the characteristics of the dataset to achieve the end-to-end classification of three types of invoice images,such as value-added tax invoices,train invoices and taxi invoices.Experimental results indicate that the proposed model achieves the best performance,and its accuracy is 99.05 %.In terms of running speed and model scale,it takes approximately 0.18 seconds to determine the category of each invoice image,and its scale is 3.65 MB,and it means that it is easier to apply the proposed model to the portable devices.Thirdly,this thesis proposes a method based on deep learning to detect and identify the invoice images taken by smart phones.First of all,YOLO-V3 model is trained to detect and extract four regions with useful texts,nearby buyer informatio n,product information,price information and seller information.Then,CTPN model is trained to extract the line text regions.Later,Dense Net model is trained to identify the line text regions and obtain the text information.The above models are trained separately,but are tested in order.Experimental results indicate that the performance of the proposed method is well.In the detection phase,the intersection-over-union of the convergent model reaches up to 1,and it takes approximately 0.66 seconds to detect one invoice image.In the identification phase,the accuracy reaches up to 95.18 %.The mothod takes approximately 1.55 seconds to identify all the characters in one invoice image,and it takes only 0.06 seconds to identify every line text region.Last but not least,this thesis summarizes and prospects the proposed methods.The feasibility of the methods is analyzed,and the possible improvements are put forward,respectively.