Method and system for generating training data

Abstract

The invention discloses a training data generation system, which includes a text generator, a layout generator, a text renderer, a chart renderer, a noise adder, a deformation adder, an annotation generator, and a document image sample library; the text renderer selects text line and render it to the position area of ​​the text line; the chart renderer renders the chart element to the position area of ​​the chart element; the markup generator generates layout analysis markup information, text positioning and identification markup information. In addition, the invention also discloses a method for generating training data. The present invention considers various factors in the document image generation process, modularizes, parameterizes, and configures the factors, and can automatically generate training samples applied to document image analysis and recognition systems, and the synthesized document images are rich in form, The effect is realistic, and the overall flexibility is adjustable, and the scalability is strong. It can also automatically complete the labeling of information at all levels of document images, and provide fully labeled training data.

Description

Technical field[0001]The present invention relates to the field of image synthesis and labeling, and in particular, to a training data generation method and system.Background technique[0002]In the prior art, the depth learning method represented by depth neural network has been widely applied to various image recognition systems. Among them, the document image analysis and recognition system uses a computer vision method to analyze the physical and logical structure of the document image, and form a complete description of the document by positioning and identifying various document elements within the document such as text, tables, images, graphics, etc. to form a complete description of the document. At the same time, training data automatic generation technology is the technique that is often used in the machine learning, and the training data automatic generation technology expands data by adding various variables to real data, or it directly generates simulation of real data mo...

AI Technical Summary

Problems solved by technology

[0003] However, the inventor found through research that in the prior art, the document image analysis and recognition system still lacks sufficient and accurately labeled training data sets due to the complicated manual labeling process of its training samples, especially the training of Chinese document images. The data set is extremely scarce; and the document image synthesis and labeling technology in the prior art only considers the individual influencing factors in the process of document image generation, and the information of the label is not comprehensive enough, which has great limitations in practical applications and cannot Applied in a complete document image analysis and recognition system

Images