Intelligent seal identification method

Abstract

The invention provides an intelligent seal identification method. The method comprises a seal area positioning process, a seal image preprocessing process, a model pre-training process and a seal text generation process. The model pre-training process is multi-modal network training, and comprises seal image model training, seal text model training and training model matching degree judgment. According to the method, on the basis of a multi-mode model of a BERT model, an image end inputs image features of backbone network coding, and a text end generates a network model through uniLM; training data are generated as supplementation through a seal image automatic construction method, and training comprises pre-training of image and text matching; during testing, recognized characters are output step by step until an end symbol by inputting a coded seal image. Compared with an image correction and general OCR Recognition method in the prior art, the method has the advantages that error accumulation is avoided, and meanwhile, the multi-modal language model has higher recognition rate and accuracy on similar characters in company names.

Description

technical field [0001] The invention relates to the technical field of graphic and text processing, in particular to an intelligent seal recognition method. Background technique [0002] In order to improve the effectiveness of seal recognition, prevent and reduce the huge losses caused by seal crimes to society, collective units, and individuals, using high-tech visual technology to identify seals accurately, quickly and efficiently has become an urgent problem to be solved; how to quickly, accurately and efficiently Distinguishing the authenticity of seals has great practical significance for individuals, social groups, and units, and has become an urgent need for social stability and development. [0003] At present, there are already some methods of using optical character recognition technology (OCR) to recognize seals in the prior art, but the following technical problems exist in the existing seal recognition methods: one is the existing pipeline method based on corre...

AI Technical Summary

Problems solved by technology

[0003] At present, there are already some methods of using optical character recognition technology (OCR) to recognize seals in the prior art, but the following technical problems exist in the existing seal recognition methods: one is the existing pipeline method based on correction and recognition, and there is a process Second, in the prior art, there is no targeted optical character recognition technology (OCR) recognizer for the stamp text, which will cause the recognition effect to be affected by similar fonts (such as: "online" The word "line" in the middle is easily mistaken for the word "money")

Images