General Document Identification System Based On OCR Technology

In recent years,with the advancement of social networking processes,the digitization process of various paper documents has also accelerated.Most of the conversion from paper documents to electronic documents is through optical character recognition(OCR)technology.After nearly ninety years of development,the research on character recognition technology has been very successful,and various OCR softwares appear on the market in an endless stream.Although the application of character recognition has made great progress,most of them are focused on document recognition,which is still a problem for document identification.Due to the complicated layout and diversified contents,the documents greatly increase their difficulty of identification.In view of this situation,we designs a universal identification system based on the original OCR technology.The general document identification system in this paper mainly includes three module preprocessing,layout analysis and OCR engine.Preprocessing includes four operations of graying,denoising,binarization,and tilt correction.Taking into account the difficulty of color image processing,the weighted average method suitable for human visual observation is used to grayscale the ID image.In order to ensure that the boundary features of the image are not blurred,the method of adaptive median filtering is used to denoise the image.The binarization of the image can highlight the target content of interest.Selecting the point-by-point method to binarize the grayscale image can make a good distinction between the foreground and the background of the ID image.Using the Hough Transform can correct tilted document images,greatly increasing the system's ability to process poor samples.The layout analysis module includes the process of template making,type judgment,attribute judgment,and area merging.In order to be able to identify the items that need to be identified in the document,a typical document template needs to be created and saved.By using a perceptual hash algorithm,the type of targetdocument entered is determined in comparison with a typical template.After the completion of the type judgment,a single item is obtained for the document cutting,and the projection principle is used to determine the item's attributes,distinguishing characters,tables,and picture elements.After the attribute judgment is completed,the complete project information is obtained by extracting and merging through the connected domain.The OCR engine adopts a deep learning method.The local convolutional neural network based on the local adjustment mode can quickly adjust the network parameters,accelerate the training process,and improve the recognition accuracy.Nearly one thousand samples of three kinds of documents including food business licenses,catering service licenses and food circulation permits were used to verify the system,and relatively good experimental results were obtained.The main information of sample documents can be accurately identified,and the ability of the system to handle complex documents is strong,reaching the level of practical application.In addition,the identification of ID cards on the general document identification system indicates that the recognition accuracy rate has reached the average level of similar documents on the market.