Research On Key Technology Of Seal In Bill Recognition Scene

The seal,as a legally effective citizen identification tool in our legal system,plays an important role in today’s economy and society,and is widely used in key locations on paper materials such as bills and contracts to ensure material authenticity.As OCR technology has continued to mature in recent years,people have gradually changed the work of electronically processing paper-based material information from manual transcription to automated processing with deep learning text detection and recognition models,which has greatly saved manpower and material resources.However,in the relevant area of the seal overprint,the text content detection and recognition accuracy rate is greatly discounted due to the mutual interference and mutual influence of the seal and the background text,which seriously interferes with the process of electronicization of paper-based information.To address this problem,this paper draws on the research results in the fields of text erasure research and image restoration to study the seal erasure method without traces: 1)to address the problem of the lack of data for seal erasure research,a variety of real data and synthetic data production methods are proposed to meet the research needs;2)a seal erasure method based on the Erase Net scene text erasure network framework is proposed to realise an end-to-end training process for the end-to-end training and processing process for stamp erasure;3)proposed a Ushaped architecture image generation network based on gate convolution,and adopted a twostage network processing idea of stamp segmentation and stamp erasure to achieve a stable,efficient and accurate final erasure effect,with a 60% improvement in text character accuracy before and after the stamp erasure processing.In the instrument,not only does the background text contain information,but the seal represents the identity of the legal entity and itself contains key information such as the executor and the contracting parties.A complete electronic system for paper-based materials will often seek to obtain the text information of the seal therein for secondary verification of the material text information and to ensure the legality of the material.To address the problem of extracting text information from seals,this paper proposes a two-stage text detection and text recognition scheme for extracting text information from seals,specifically: 1)Considering the problems of diverse and complex seal styles and random text alignment directions,a multi-scale seal text region segmentation detection network based on image pyramids is proposed,which solves the problems of difficult detection of curved text and irregular text region shapes of seals.The final detection F1 score reaches 92.24%.2)To address the problem that the angle,shape and text arrangement of text blocks are not fixed after seal text detection,a set of text correction methods based on image classification and conversion of image polar coordinates to rectangular coordinates are proposed,and the design of CRNN backbone network and semantic feature extraction network is optimized to realize the extraction process of seal text information,and the final text recognition The accuracy rate reached94.90%.