Research On Automatic Identification Technology Of Business License

With the rapid development of artificial intelligence,the scene text recognition method based on deep learning is accepted and used by more and more people,and the recognition effect is becoming more and more prominent,which has gradually become one of the hot research contents in the computer field.Due to the complex background of the image,there is a lot of useless information in the image features,resulting in the low Recognition rate of traditional OCR(Optical Character Recognition),which cannot meet the needs of customers.Therefore,this paper USES convolutional neural network to realize the detection and Recognition of image characters.As the authentication module of the major science and technology special project of xinjiang uygur autonomous region,"construction of technological innovation platform for horse industry",Need to authorize the authentication of the unit,this study was carried out according to the requirements of this project.In this paper,the latest version of the business license is taken as the research object.However,there are some problems in the image of the business license,such as uneven illumination,fuzzy motion,and complex background.Therefore,the main research content of this paper is to accurately identify the target text in the image of the business license through convolutional neural network.Currently on the background of complex image character recognition technology is mature,the id card,train ticket,the respect such as bank bills,invoice is widely applied,but for business license character recognition research is few,open source identification model of character recognition is low,directly used in project is not feasible for certain,Therefore,the existing model is optimized and retrained to improve the text recognition rate of business license.There are many methods about deep learning to realize text recognition,and each algorithm has certain advantages and disadvantages.This paper focuses on the current popular text detection and text recognition algorithms,which are CTPN(Connectionist Text Proposal Network),CRNN(Convolutional Recurrent Neural Network)and DenseNet(Dense Convolutional Network).The text recognition models based on CTPN+CRNN and CTPN+DenseNet are designed to realize end-to-end image text recognition.Both models have been encapsulated and can be switched for use in the Demo according to the different data used.At first,This paper with the manual annotation of 2500 Zhang business license to the open source CTPN network training again,the AP value of CTPN reached 94%.after tests found CTPN font small text will appear in the business license the testers,and when a multi-line text each line has not been separated detection,in order to solve this problem,the business license for the layout analysis,and regional segmentation,get 10 sub images,it is good to solve the above problems,and the AP value of text detection after the above process reached 98%,the AP value of CTPN increased by 4%after treatment.Secondly,541126 variable length character data were used to retrain the CRNN model.Finally,it was determined that the more suitable recognition model for business license was based on the CTPN+CRNN text recognition model,and the final text recognition rate reached 96%,which has certain reference significance for the method of business license text recognition.