Hamming OCR method based on locality sensitive Hash network

Abstract

The invention discloses a Hamming OCR method based on a locality sensitive Hash network. The method comprises the following steps of: S1, acquiring a scene text image with OCR for making a training set and a test set; s2, constructing an auxiliary model, optimizing the auxiliary model by using the text image of the training set, training the model, and generating locality sensitive hash codes of all characters; s3, constructing a Hamming OCR master model, and outputting the locality sensitive hash code as an embedded layer of the Hamming OCR master model to obtain an initialized Hamming OCR master model; s4, optimizing parameters of the initialized Hamming OCR main model by using the text image of the training set to obtain an optimal Hamming OCR network model; and S5, inputting a to-be-recognized text image into the optimal Hamming OCR network model, generating an output code by the optimal Hamming OCR network model, and obtaining a text recognition result through the output code. According to the method, the model storage occupation capacity is reduced, and the model stability and the operation efficiency are improved.

Description

technical field [0001] The invention relates to the field of text recognition, in particular to a Hamming OCR recognition method based on a locally sensitive hash network. Background technique [0002] Scene text recognition aims to extract text content from images, which has great commercial value in practical applications. Scene text recognition usually includes scene text positioning and text recognition. For example, in Chinese text recognition, there are already a large number of mature detection and recognition methods at home and abroad, and they have been applied to practical work. These methods not only detect quickly, but also have a good recognition rate, which has certain advantages. Although scene text recognition has achieved certain breakthroughs and produced many representative models, most of these models are cumbersome. The general scene text recognition model uses one-hot encoding to encode each character in the final regression classification. When the...

AI Technical Summary

Problems solved by technology

Although scene text recognition has achieved certain breakthroughs and produced many representative models, most of these models are cumbersome.

The general scene text recognition model uses one-hot encoding to encode each character in the final regression classification. When the vocabulary is large, the mapping matrix generated by the character embedding layer and regression classification layer of the model will occupy a large amount of storage capacity. These models are difficult to deploy on mobile devices because it consumes a lot of storage resources and computing space

Images