Invoice Recognition System Based On Deep Learning

The current invoice reimbursement work is still based on manual reimbursement,which has many disadvantages such as low efficiency,lasting reimbursement process,excessive time consumption,and high error rate.At the same time,the growth of economy has also led to an increase in bills and reimbursement.Therefore,it is a general trend to achieve the intelligent identification of invoices.In view of this topic,this paper mainly explains how to implement invoice intelligent identification from two aspects of invoice classification and text recognition.First of all,this paper briefly introduces the current status of text recognition and Deep Learning.After the description of Deep Neural Network,it builds an AlexNet network and trains the model with the data set of invoice to achieve the purpose of classification,then uses the networks of AlexNet and Random Forest to test this model respectively.Secondly,this paper adopts a new method of completing the positioning of effective information in the invoice and correcting the oblique image.The method of quadratic segmentation based on image enhancement is used to implement the effective information location and interception of invoices.In the first step,the image is enhanced to facilitate the subsequent segmentation of machine-printing characters and typographic characters,which is the so-called 'the first segmentation'.The second step is to use the inverse color,open operation,maximum rectangular boundary and other image processing methods for the acquisition of information,which is the so-called 'the second division'.The skew detection and correction algorithm is used to detect the tilt angle of the capital amount character and correct the tilt image.Finally,this paper builds the architectures of Deep Convolutional Neural Network and Residual Network,which are used to train and test the models with the capital amount price characters separately,and finally realizes the identification.In this process,it is also necessary to use the transformation of projection on the pre-processed capital amount to cut the line into a single character,which is the production of the data set.In conclusion,the invoice classification accuracy of the AlexNet network and the network of Random Forest is 92% and 94%,moreover,the recognition accuracy of Deep Convolutional Neural Network and Residual Network is 97% and 99% respectively.The method of intelligent identification in this paper can be applied not only to invoice reimbursement,but also to other fields,such as file digitization.