Design And Implementation Of A Bill Recognition System Based On Deep Learning

With the rapid development of deep learning technology and the widespread application of convolutional neural networks in computer vision research in recent years,research work in the field of image analysis and understanding has achieved remarkable results.As an important branch of image analysis and understanding,Optical Character Recognition(OCR)technology has gradually matured and is widely used in life.The demand for OCR technology for document recognition and scene text recognition is growing day by day,so it is of great use value and application prospects.Bills,as a paper document that people often use in their lives,record important consumption information.However,in the process of bill information review and financial reimbursement,handling a large number of bills is a tedious task and in the process of information upgrading in the financial company,the processing and filing of paper bills is also time-consuming and labor-intensive.Thanks to the wide application of OCR technology in the field of document recognition,some researchers have studied using OCR technology to automate the processing of bill pictures,locating the text on bill pictures and identifying text information.However,the text target on the bill picture is extremely dense and due to factors such as printing quality,paper damage,handwritten signature,scanning quality and other factors,it is difficult to recognize the bill text.In addition,there are fewer bill recognition systems for users,and the recognition accuracy is lower,which cannot meet the actual application requirements.Therefore,researching text recognition algorithms for bill pictures and establishing bill text recognition systems are of great research significance and value.This paper starts with the construction of a bill text recognition system,then studies the existing problems and technical difficulties of OCR technology in bill text recognition tasks,and designs text detection and recognition algorithms according to the characteristics and technical difficulties of bill text.Aiming at the problem that the segmentation algorithm has adhesion on dense note text,this paper designs a text detection algorithm based on modeling the boundary area between adjacent text units,proposes a boundary modeling idea and an alignment loss function design to solve the problem of adhesion of dense text segmentation area,introduces differential binarization module to solves the problem of insufficient training of the segmentation network and improves the accuracy of note text detection.Aiming at the problem of low recognition accuracy of mainstream text recognition algorithms on bill data,on the basis of analyzing the characteristics and method shortcomings of bill text,the improved design is proposed from three aspects of training method,network structure,and post-processing to improve the recognition accuracy of bill text,which provides text detection and recognition algorithms with high accuracy and high robustness for the bill recognition system based on deep learning.In order to carry out research on bill text detection and recognition algorithms and verify method performance,this paper proposes a bill text data set(BTD).Finally,on the basis of the algorithm designed in this paper,through the demand analysis and functional analysis of the bill recognition system,the web-based deep learning bill recognition system is realized to meet the actual application needs.