Research On Business Card Recognition Based On Convolutional Neural Network

Optical Character Recognition(OCR)is widely used in document recognition.Useful information can be quickly extracted from them by digitizing pictures of characters.Traditional optical character recognition algorithms mostly depend on manual design features.Template matching method is used to realize the recognition of specific scenes.Therefore,the applicable scenario is relatively single,the generalization ability is poor,and the effect is not good when processing the task of business card recognition.In addition,the traditional convolutional recurrent neural network(CRNN)is not ideal for text image detection with noise interference between words.Therefore,in view of the above problems,the text focuses on OCR business card recognition based on deep learning to make up for the shortcomings of the traditional recognition system.From the perspective of information extraction,this paper uses OCR technology to identify business card information,and then computerizes business card information to realize structured storage of business card data.Based on the analysis of traditional methods and current mainstream methods,the text has been improved and optimized appropriately,and a new OCR recognition system based on convolution neural network has been realized.In the aspect of image preprocessing,this paper designs a set of preprocessing flow for business card image,such as edge detection,tilt correction,etc.to eliminate the influence of image interference factors.In addition,aiming at the image blur caused by camera jitter,this paper proposes and implements a deblur model based on encoder/decoder network to improve the effect of subsequent character recognition.In terms of text area detection,this paper proposes and implements a text area detection method for business card recognition.Based on the Yolo network,the final detection accuracy is improved by 0.6% by using the fixed width text image.In terms of text recognition,in order to improve the recognition rate under the condition of mixed Chinese and English,targeted training is carried out to improve the accuracy of text recognition by 1.6%.Finally,in terms of system implementation,the system's humancomputer interaction mode was designed,using the B / S architecture,and using an efficient Flask framework on the front-end Web server.On the back-end server,the OCR processes are modularly designed,and finally the structured output results are returned.