Discrete picture file information extraction system and method based on deep learning

Abstract

The invention discloses a discrete picture file information extraction system and method based on deep learning, an end-to-end OCR model combines text detection with text recognition, the sharing of an image processing layer of a convolutional neural network is achieved, and the calculation speed and recognition effect are improved. Two-way LSTM is used in the aspect of character detection, a textarea is accurately segmented according to file layout when characters are recognized, and the problems that noise can be recognized as the characters through an existing OCR technology, messy codes can be generated for a chart, and wrong lines can be generated can be solved. In the aspect of text recognition, a ResNet is combined with a Transformer codec, so that written bodies and handwritten bodies can be accurately recognized; by combining an end-to-end OCR model with an information extraction model, information loss can be reduced, and the accuracy of information extraction can be improved; and the Bayesian optimization algorithm is used to realize training and parameter adjustment automation of the information extraction model, so that a user can customize the desired information extraction model even if the user does not understand machine learning.

Description

technical field [0001] The invention relates to the technical fields of artificial intelligence, optical recognition and machine reading lights, and in particular to a system and method for extracting discrete image file information based on deep learning. Background technique [0002] Optical Character Recognition (OCR) refers to the process of analyzing and recognizing image files of text materials to obtain text and layout information. That is to recognize the text in the image and return it in the form of text. [0003] Existing OCR technologies such as figure 1 As shown, it mainly includes image preprocessing, text detection and text recognition. Image preprocessing usually corrects the imaging problem of the image. Common preprocessing processes include: geometric transformation (perspective, distortion, rotation, etc.), distortion correction, blur removal, image enhancement, and light correction. Text detection is to detect the location, range and layout of text, a...

AI Technical Summary

Problems solved by technology

[0014] However, the FOTS model has some problems in extracting sparse text: 1. The algorithm mainly relies on CNN for detection, and the LSTM encoder is only used for recognition. It is difficult to distinguish printing and handwriting on the same line when the size of the training set is limited. Text; 2. The application scenario of FOTS is text detection and text recognition of natural pictures, rather than scanned documents. Using the EAST algorithm that can detect text at any angle, the accuracy of long text detection will decrease; 3. Due to the use of LSTM+ CTC, its recognition algorithm will decrease the recognition rate for noisy pictures, which is not robust enough

The existing OCR technology mainly has the following problems: 1. Users need different information for different texts, and a non-customized pervasive system cannot accurately extract the required information. For example, for employee information forms, users may need Extract the employer and employee's name, salary, position, etc. For invoices, the user may need to issue the invoice's unit and address, product name, price, and tax amount in advance

2. The current OCR technology does not have the ability to extract handwritten characters. The current handwritten font recognition based on deep learning can only deal with simple text, such as a line or a paragraph of dense handwritten text, but the text in the real environment often needs to deal with complex tables. or handwritten fill-in-the-blank signatures, etc., and handwritten information is generally important information that needs to be extracted. At present, there is no mature solution to locate, classify, and identify

3. In most current OCR algorithms, text detection and text recognition are separated, and the inaccuracy in detection and inaccuracy in recognition are superimposed, resulting in a decrease in the overall recognition rate

4. The current end-to-end technical solutions are limited to character extraction or ordinary handwriting recognition, which requires that the text provided by the user has been pre-processed, that is to say, what the user sends to the system is the image information that the user wants to extract. For example, the user If they want to import salary data from 5,000 scanned employee forms of different companies into the database, they need to extract part of the salary pictures from the 5,000 scanned pictures, and then send them to the OCR system, which is obviously very unrealistic

Images