Data processing method and device

By using the attention interaction between the text recognition network and the text detection network in end-to-end text recognition, the problem that feature correlation and differences in the prior art are not effectively utilized, and the accuracy of text recognition and the optimization of model structure are improved.

CN119723583APending Publication Date: 2025-03-28HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311287808.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-09-28
Publication Date
2025-03-28

AI Technical Summary

Technical Problem

The existing end-to-end text recognizer based on Transformer failed to effectively distinguish the feature correlation and differences between different task branches, resulting in redundancy in model structure or feature confusion, limiting the performance of text recognition.

Method used

By acquiring images and using the text recognition network to obtain feature representations of the image and the center position recognition results of the text box, and using the text detection network to perform attention interaction, we obtain the edge detection results of the text box, so that the recognition feature guides the detection task and optimizes the model structure.

Benefits of technology

It effectively avoids the accumulation of detection errors in the traditional detection first and then recognition framework, and uses the correlation and differences between detection and recognition features to improve the accuracy of end-to-end text recognition, and avoids the problems of redundant structural design and feature confusion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119723583A_ABST
    Figure CN119723583A_ABST
Patent Text Reader

Abstract

The invention discloses a data processing method. The method comprises the steps of obtaining an image; obtaining a first feature representation of the image and a central position identification result of a textbox in the image through a text identification network; and performing attention interaction on the first feature representation and a second feature representation corresponding to the central position identification result through a text detection network to obtain an edge detection result of a textbox in the image. According to the method, the recognition features and the recognition result are used as priori of the detection task, and text detection is performed by using the output of the text recognition network, so that the accumulation of detection errors in the recognition stage can be effectively avoided, and the correlation and difference between the detection features and the recognition features can be effectively utilized; therefore, the end-to-end text recognition precision is improved.
Need to check novelty before this filing date? Find Prior Art