Research On Automatic Identification System Of VAT Invoice Content

VAT Invoices play an extremely important role in social and economic activities,especially since there is a lot of important information on invoices that needs to be stored on computers.In order to reduce the cost of manual entry,this paper designs an automatic identification system of VAT invoice based on convolutional neural network.The system is an automatic system based on convolutional neural network,which contains two different convolutional neural networks,one for detecting the target region and the other for obtaining the specific character information of the target region.The system uses a scanner to collect multiple images of invoices.The system is mainly composed of 6modules: input,preprocessing,segmentation,recognition,location and recording.The key modules of the VAT invoice identification system are positioning and recognition.The accuracy and speed of these two modules determine the quality of the whole system.In order to solve the problem that the traditional positioning method cannot balance the detection accuracy and time delay well,YOLOv3 detection algorithm is introduced into the positioning module of the system to dectect the target area of invoice,so that the detection accuracy can be better improved without too much time delay.In the aspect of character recognition,template matching has its own limitations.Once the matching object in the original image has rotation or difference size from template,the limitations of template matching determine that the algorithm cannot achieve accurate matching in the case of multi-scale or slightly distorted graphics.In the recognition module of the invoice system,in order to solve the limitation of template matching,the Inception convolutional neural network is used to recognize the characters,so that the image can be accurately recognized in general.The main contents and achievements of this paper are as follows:1.Improvement of image preprocessing moduleIn this paper,in the image preprocessing module of the automatic identification system of VAT invoice contents,the judgment condition of invoice inversion is improved to solve the problem that the judgment of invoice inversion is prone to misjudgment.Inthe segmentation module of the automatic identification system of VAT invoice contents,OTSU algorithm is used to obtain the adaptive binarization threshold of the image in order to solve the problem that the foreground and background are not accurately separated by manual binarization threshold.In order to solve the problem of partial smudge or uneven brightness of the invoice,the image is divided into several blocks using the OTSU algorithm.2.Location of target region based on convolutional neural networkWith the rising improvement of deep learning and convolutional neural network technology,many computer vision tasks performance has gone beyond the traditional method.Therefore,the invoice system is introducted conventional neural network model to replace the traditional algorithm for target localization tasks.In order to be better than traditional algorithm in the detection for a balance between speed and accuracy,and to be compared with traditional image matching algorithm has better robustness,YOLOv3 and its some basic Settings is adjusted for VAT invoice target location.Finally,the YOLOv3 algorithm is compared with the traditional SVM + HOG algorithm for the speed in target detection.3.Character recognition based on convolutional neural networkDue to the particularity of Chinese characters,most of the early invoice identification systems have solved the problem of digital recognition well and achieved high accuracy,but they have not effectively solved the problem of Chinese character recognition.In the past,Chinese character recognition used some artificial rules to complete the recognition matching,such as feature point matching,pixel matching,computing the Euclidean distance and cosine distance of image matching.Using these matching methods can degrade recognition performance as the quantity of templates and types increases.In order to overcome the difficulties of matching failure caused by noise disturbance,half-occlusion,ambiguity,scale inconsistency or improper angle in template matching,the convolutional neural network in deep learning technology is used for character recognition in the recognition module of the system,so that the image can be accurately recognized in general.