Multilingual end-to-end OCR algorithm and system

Abstract

The invention provides a multilingual end-to-end OCR algorithm and a multilingual end-to-end OCR system, which overcome and bypass the defects of fragments in the prior art and have excellent performance on character adhesion, Chinese-English and median mixed data. By means of a self-distillation transformer module, the position relation is reserved, parameters and model complexity are reduced, meanwhile, results are output in parallel, dependence between nodes is cut off, higher robustness is achieved for multilingual and multi-font scenes, and the structure and performance are optimized. Thealgorithm comprises the steps of obtaining a feature map of a to-be-identified picture; training the feature map through a relation attention module based on the self-distillation transformer moduleto obtain a character matrix; performing parallel attention decoding on the character matrix to obtain a prediction result; and according to the prediction result, based on the vocabulary sentence table, obtaining an OCR model conforming to the language of the vocabulary sentence table.

Description

technical field [0001] The invention relates to an OCR algorithm, in particular to a multilingual end-to-end OCR algorithm and system. Background technique [0002] In the modern enterprise production environment and daily business activities, OCR—Optical Character Recognition: optical text recognition technology has been fully promoted and applied on a large scale thanks to the increasing demand for improving information entry efficiency and diversifying information carriers. In existing industrial application scenarios, considering compatibility and reliability, OCR applications are usually divided into two parts: text detection and text recognition. The text recognition network is also divided into a single word classification recognition network structure and an overall recognition network structure. The existing overall recognition network usually adopts CRNN (Convolutional Recurrent Neural Network) structure, which has the following disadvantages: [0003] 1. Usually...

AI Technical Summary

Problems solved by technology

[0003] 1. Usually the input text image is a feature sequence with spatial or plane information association. Using the existing overall recognition network, which needs to be reshaped by the Recurrent layer, the context information constructed by it will inevitably be lost, especially the spatial position. information

However, if the 2D-Recurrent layer is used for encoding to solve the problem of information loss, it will make too many parameters and complex structure

[0004] 2. Use the existing overall recognition network. Generally, LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) networks have requirements for the width of input slices, and in text information, there are often mixed Chinese, English, and Chinese numbers. Sticking, etc., resulting in different character widths, and the adaptive width fragmentation structure is difficult to have better robustness in these scenarios

[0005] 3. For the OCR pipeline of positioning and recognition classification, it is usually difficult to achieve the optimal anchor box size selection, which requires lengthy iteration and labeling process

Images