Research And Implementation Of Character Recognition System Based On Deep Learning

With the rapid development of computer technology,as an important communication carrier in information technology,image occupies a very important position in the dissemination of information.In the era of big data,how to efficiently process massive image information has become a hot research issue^[1].With the surge of artificial intelligence,it has become a trend to use highly intelligent machines to replace manual work in all walks of life.Optical character recognition(OCR,optical character recognition)technology is an important part of computer vision.It simulates human vision to intelligently recognize and judge the information in the image.OCR technology aims to detect and identify text information from pictures,that is,the process of scanning text data,and then analyzing and processing image files to obtain text and layout information.It is mainly used for document identification and identification of documents,through automated processing,to reduce labor costs,improve work efficiency,and reduce the error rate.This paper uses deep convolutional neural network to implement a character recognition system suitable for bill documents.Bill documents include ID cards,business licenses,retail licenses,value-added tax invoices,transaction confirmations,etc.The input image is taken by a camera device such as a mobile phone or a camera.The system designed in this paper combines digital image processing,deep learning and natural language processing technologies to complete the Chinese OCR recognition process for specific application scenarios.The main work of this article is as follows,1.Aiming at the problem that the image quality of the document bills taken by mobile phones is uneven,and the images are likely to contain redundant background information,this paper proposes an algorithm to extract the region of interest(ROI)based on the target background color.Pre-processing,locate the specific location of the document bill in the image,remove the interference of the background information in the image,and improve the recognition effect of the entire character recognition system.The algorithm is divided into three background colors of green,blue and red according to the different background colors of the document bills contained in the image.The edge detection and morphological processing of the image are used to obtain the target(document bill)in the entire image Location coordinates.Experiments show that the preprocessing process of the image region of interest(ROI)extraction algorithm based on the target background color can remove the interference information well and improve the OCR recognition effect.2.Investigate and analyze the current character recognition algorithm.The current mainstream character recognition algorithm framework is"feature extraction network(CNN)+circular convolution network(RNN)+CTC(Connectionist temporal classification)algorithm".In this paper,the mainstream character recognition algorithms using Google Net,Res Net and Dense Net as feature extraction networks are tested,and the effects of the algorithms are compared and analyzed.In order to solve the problems that the mainstream character recognition algorithms require too high hardware conditions,occupy too much memory,and the calculation speed cannot meet the real-time requirements,two lightweight character recognition models are proposed.One is an improved Dense Net lightweight character recognition model,and the other is a lightweight character recognition model based on deep separable convolution,and the two lightweight networks are combined with the existing mainstream character recognition network.A comparative analysis was carried out.Experiments show that the two lightweight character recognition algorithms proposed in this paper have smaller models and faster calculation speeds than traditional character recognition algorithms.In particular,the character recognition network based on deep separable convolution benefits from its different convolution methods,and the algorithm performance is more excellent.3.Because the character recognition process uses convolutional network for recognition,it cannot achieve 100%accuracy.In view of the error recognition problem in the recognition results of the character recognition model,this paper proposes a Chinese shape near word error correction algorithm based on natural language processing.The back end of the character recognition network further improves the detection accuracy of the character recognition system.The character recognition network judges and outputs according to the morphological characteristics of the text.The characters that recognize errors and the correct characters are mostly similar in morphological characteristics.Therefore,the error correction algorithm proposed in this paper is mainly for errors between Chinese near-characters.This paper proposes two error correction algorithms for Chinese near-characters.One is an error correction algorithm based on Hidden Markov Model(HMM).It is after detecting a sentence error,Use the font to replace the detected errors to complete error correction.The other is an error correction model based on the encoder-decoder mechanism,which performs error correction based on the semantic information provided by the context.Through experimental comparison,the error correction algorithm based on HMM has a faster error correction speed,and the error correction algorithm based on the encoder-decoder mechanism performs well in the long-sentence error correction scenario,both of which can serve well for the identification and correction of document tickets.4.Integrate the algorithm proposed in this paper to design a character recognition system for document bills,and use the pyqt5 tool to complete the design of the interactive interface between the system and the user.This system realizes the process of character recognition of the document bill image obtained by the mobile phone,camera and other camera equipment,which is converted into text information.It is mainly divided into three parts,including the extraction of the image target area,character recognition,and error correction.Users can create a custom dictionary in the interactive interface according to their own needs,and select the background color of the document bill to be identified.After the recognition is completed,the user can get the recognition result of the character recognition network and the final output result after error correction.After testing,in the application scenarios set in this article,the recognition accuracy rate can reach 98.37%.The character recognition system designed in this paper can well meet the accuracy requirements for character recognition of documents and bills.At the same time,due to the introduction of a lightweight network,the system can also meet the real-time recognition needs while ensuring recognition accuracy.