Research On Credential Image Text Detection And Recognition Based On EAST And CRNN

With the rapid development of digital information technology,intelligent identity authentication has been widely used in many Internet companies and government departments.If there is a method that can automatically and accurately identify various key text information in the credential image taken by users,the tedious manual process will be simplified and the work efficiency can be improved greatly.Traditional Optical Character Recognition(OCR)technology mainly aims at the recognition of printed and scanned document image,but there still exists challenges when recognizing credential image which requires high recognition accuracy.In recent years,deep learning has made great achievements in various tasks in the field of computer vision.This paper is based on an Efficient and Accurate Scene Text detector(EAST)and a Convolutional Recurrent Neural Network(CRNN)which can recognize various length of text sequence end-to-end in deep learning era,and improves the two methods according to the characteristics of credential image text with good results.The specific work is as follows:This paper proposes a credential image text detection method based on positionsensitive regression to alleviate the limitation of receptive field in EAST.The proposed method predicts the coordinate offset between each side text pixel to its corresponding vertexes of bounding box instead of the distance between each text pixel to the four bounding box edges in EAST.The predicted coordinate offset combines the predicted text score map and the text box reconstruction module proposed in post processing will form the final detection result.The proposed method achieves an F-score of 91.3%on the credential image dataset and has good competitiveness on the scene text detection datasets.According to the fact that there is no mechanism in CRNN to model the dependence between each predicted character,this paper designs a credential image text recognition method combining Connectionist Temporal Classification(CTC)and attention mechanism to decode.The designed method uses an encoder based on global context block,which can effectively fuse the global context features of the input image,and combines CTC and attention mechanism to decode,which can incorporate the language modeling to the recognition model and make use of strong context correlation of the text instance in the credential image to improve the recognition accuracy.The designed method achieves an accuracy of 96.5%on the credential image dataset and has good performance on several scene text recognition datasets.