Form information extraction system and method with privacy protection

Abstract

A form information extraction system and method with privacy protection, comprising: a node unit feature collection module, a user privacy-sensitive desensitization module and a neural network collection module located at a local end, and a graph neural network module located at a cloud service end, node unit feature collection The module identifies the node according to the image to be recognized by the user input, and obtains the text field and coordinate features of the node through the text recognition model and text positioning model deployed on the local end. The user privacy-sensitive desensitization module uses the self-attention mechanism to convert text through the coordinate space. The spatial transformation transforms the text and coordinate information of each node and obtains the vectors of the original semantics in different dimensions. The neural network acquisition module extracts the image features through the convolution operation of the convolutional neural network, and the graph neural network module extracts the image features. According to the node characteristics of the node connection relationship, the graph convolutional neural network is used to understand the location characteristics and adjacency relationship of the learning nodes, and finally the node connection relationship is obtained. The node text and coordinate information obtained by the node unit feature collection module and the node connection returned by the cloud server relationship, extract the entire picture information and restore the entire table.

Description

technical field [0001] The invention relates to a technology in the field of artificial intelligence applications, in particular to a system and method for extracting form information with privacy protection. Background technique [0002] In the prior art, for unstructured table documents, the automatic extraction accuracy needs to be further improved, and usually requires manual processing after processing. This often becomes the bottleneck of the system processing speed. With the rise of data mining and machine learning technologies, deep learning has been widely used in many aspects. Similarly, in the field of table recognition, existing neural networks are used for table recognition and table recovery. However, in the identification process in the prior art, information is often lost due to the defects of traditional neural network technology. At the same time, most of the existing table recognition does not have the ability to protect privacy, it is difficult for ordi...

AI Technical Summary

Problems solved by technology

However, in the recognition process of existing technologies, information is often lost due to the defects of traditional neural network technology.

At the same time, most of the existing table recognition does not have the ability to protect privacy. It is difficult for ordinary small companies to deploy a local recognition model. For the deployment of C / S mode, privacy protection is extremely important in some scenarios, and information security Gradually paying attention to today, a picture is directly passed in for identification, which is extremely sensitive to important data

Images