Research And Application On Text Detection And Recognition Algorithm For Bill Recognition

Bills have been an important voucher for recording transaction information since ancient times.In recent years,with the increasing volume of human transactions,especially the rapid development of China’s economy,the classification and collation of various types of bills has become an important work of ticket departments.To alleviate the labor pressure of ticket departments,relevant research on extracting key information from tickets through text detection and recognition algorithms has been put on the agenda.However,due to the variety of tickets,as well as the blurry printing and seal coverage of some tickets,there are still considerable challenges in extracting key information from tickets.At present,in-depth learning methods have been widely used in scene text recognition.Research on text detection and recognition related algorithms has also evolved from the traditional research based on segmentation,structure,connectivity components to the field of deep learning,and further evolved into end-to-end text recognition and text location detection,then recognition.The method of detecting the position of text before recognizing it in pictures is also called two-step method.Based on two-step method,this thesis will explore the application of the model in ticket picture recognition tasks by building a text detection model and a text recognition model.In addition,on the basis of the general model,this thesis constructs about 400,000 text bars through manual labeling and data enhancement to support the training of Chinese character recognition model in ticket detection.The main work of this thesis is as follows:First,based on the twostep method,a bill text recognition model is proposed,which combines the Progressive Scale Expansion Network(PSENet)and the Convolutional Recurrent Neural Network.By using the progressive scale expansion network in the text position detection stage,the selected text can be more accurate and more conducive to the recognition of Convolutional Recurrent Neural Network(CRNN).Secondly,on this basis,PSENet’s model and parameters are optimized and debugged for different tickets,so that it can be better applied to ticket picture recognition tasks.In addition,CRNN is also optimized for ticket text recognition,and a ticket picture recognition dataset is established.CRNN is trained on this dataset to get a CRNN model that can recognize more uncommon words and recognize more accurate ticket text.Finally,this thesis compares the performance of common algorithms for text detection in tickets,evaluates the feasibility and superiority of the algorithm,and finally forms a system to be applied to the actual text detection work.The validity and universality of the algorithm are further verified.