Research On Invoice Recognition Based On Deep Learning

With the rapid development of national economy and science and technology,VAT invoice has fully participated in some economic and trade activities in the society.The processing and filing of VAT invoice is a unified and repeated work for financial personnel of companies or organs.In recent years,with the development of neural network,in-depth learning has become a very important tool in text recognition The hot research direction,which uses the fast image processing and feature extraction technology,can effectively locate the invoice image information area and recognize the text,which is of great significance to reduce the cost of human and material resources..This paper mainly studies the text recognition function of VAT invoice.Most of the invoices collected by scanners have some problems,such as unclear handwriting,too close space between text lines,unclear character features,and different degrees of influence of seals,noise,etc.in addition,the invoices have tilt or uneven illumination in the collection process.The general scene target detection network can accurately locate the up and down of the text,but in the dense text location,there are overlapping and dislocation,which will also cause great difficulties to the text recognition.In text recognition,the traditional general data set can not be close to the real invoice text completely,so it will also cause interference to the accuracy of location.The main work of this paper is as follows:1 When using scanner to collect invoice image in real time,it will inevitably be affected by different background brightness or angle deviation during the collection process.Therefore,it is necessary to pre process the image after the collection,including binarization,tilt correction and so on.Because the overall layout of the invoice presents a standard tabular form,the table frame detection and output coordinates are added in the preprocessing to serve the subsequent positioning work.This paper presents a method of automatically estimating the average character height based on the calculation of the height of the connecting elements of the surrounding rectangle.This method,combined with the morphological operation in opencv,can well detect the table frame lines of the invoice image,and output the frame coordinates,so as to provide services for the subsequent positioning work Location of areas to be identified2 In view of the problems of poor text features,different fonts,different sizes and inter line distribution in most of the bill images collected by scanners,a text location method based on improved ctpn is proposed.This method not only combines the classification and regression part of the common loss function,but also uses the table frame coordinates obtained from the preprocessing part to add the coordinate difference in the loss calculation Confidence considerations.3 To solve the problem that the general text data set is not real,this paper uses the real invoice data set used in the real scene.At the same time,the font,size and text box size are unified.A text recognition framework of invoice based on tensorflow is designed.The recognition framework combines with convolutional neural network densenet to train character data set.Finally,we use the recognition model to recognize the invoice text.