Document Image Recognition Method and System

Abstract

The present invention provides a method for identifying a document image, which includes: inputting an image of a document to be identified; obtaining static features of the image of the document to be identified by a first neural network; transforming the static features according to a predetermined sequence of layout structures converting to a time series signal; the second neural network iteratively processes the time series signal to obtain the dynamic features of the document image to be recognized; and outputs the recognition result of the document image to be recognized according to the dynamic feature. Correspondingly, the present invention also provides a document image recognition system. The implementation of the invention can improve the accuracy and efficiency of document image recognition.

Description

technical field [0001] The present invention relates to deep learning and neural network technology, in particular to a document image recognition method and system. Background technique [0002] With the development of office automation technology, in more and more scenarios, people expect paper documents to be converted into electronic images that are easy to process, so as to facilitate data transmission, distribution, archiving and viewing. [0003] Due to the complexity of various businesses, the corresponding paper documents are different. When storing paper documents, people can classify them in the form of manual packaging for easy reference. The most common way to generate an electronic image of a paper document in the prior art is to scan the paper document. However, after paper documents are transformed into document images, how to automatically classify, archive and distribute various unmarked document images is a more difficult problem. If you simply rely on t...

AI Technical Summary

Problems solved by technology

Therefore, the defects of this prior art are: first, whether the method of feature extraction is effective often depends on the experience and knowledge accumulation of designers, and there is no universal and efficient feature extraction method that can adapt to different types of document images; For new types of document images, the feature extraction method needs to be re-determined, and redesigning the feature extraction method is time-consuming and labor-intensive, that is, the classification model does not have machine learning capabilities and cannot cope with the expansion requirements of document image recognition

However, the pure static network composed of CNN+classifier network cannot use the different layout context structures of document images to focus on the information used for classification, that is, the features in the static network are all instant and one-off, such as carrying The content of multiple document images of different tables, the difference between them is only reflected in the table structure in different positions in the image, and the static classification network composed of CNN+ classifier will not pay attention to this, so the document image in the document image The context structure information of the layout cannot be effectively utilized, which leads to the fact that the efficiency and accuracy of the recognition of the static classification network composed of CNN+ classifiers when classifying document images cannot reach the expected value

Images