Document image classification method and device

Abstract

The invention discloses a document image classification method and device, and belongs to the field of computer vision. The method comprises: training a text feature vector extraction model and an image feature vector extraction model respectively, extracting a fusion feature vector of a document image in an embedded feature mode that text feature vectors and image feature vectors are fused, and classifying the document image based on the similarity of the fusion feature vector. According to the method, various document images can be quickly registered and classified, the service process can be greatly simplified, the OCR API can be simplified, all documents can be recognized through one API, and permanent use of one-time access is truly achieved.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a document image classification method and device. Background technique [0002] In all walks of life, there are still many paper documents that need to be saved, processed, retrieved, etc., especially in the financial field such as banking, securities, insurance, mutual funds, finance, taxation and other industries. In the past, the digitization of these paper documents was generally manually entered. With the continuous popularization of OCR technology, many industries have gradually adopted OCR recognition technology instead of manual entry, which has greatly improved work efficiency. But at present, the prerequisite for good OCR recognition and structuring is that the category of the document needs to be clearly known, otherwise it is difficult to have a good structured result. In addition, in many occasions such as bank counters, the current application is that the user must ...

AI Technical Summary

Problems solved by technology

[0004] 1. Difficult to obtain many styles: There are too many types of document images, and different document types in different fields. It is impossible to collect all of them for training, and sometimes they are added later and cannot be obtained in advance. Some documents are confidential Yes, unable to train without desensitization

[0005] 2. Complicated collection methods: With the popularization of collection devices such as mobile phones, tablets, cameras, scanners, and cameras, especially mobile phones, document image acquisition methods have shifted from traditional scanning methods to shooting methods. At present, more than 90% of documents The images are taken rather than scanned. Due to the complex background of the taken images, compared with scanners, they are not as good as scanners in various conditions such as background, resolution, orientation, illumination, font, character size, etc., and cannot Uniform specification

Images