Text detection model training method, text detection method, device and equipment

CN119625704APending Publication Date: 2025-03-14GUANGZHOU SHIYUAN ELECTRONICS CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311175255.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-09-12
Publication Date
2025-03-14

AI Technical Summary

Technical Problem

The lightweight text detection model affects the accuracy of text detection due to fewer parameters, and has poor knowledge distillation effect, poor migration, high data labeling cost, and uneven distillation effect.

Method used

By fusing the target sample image and the target image, the fused sample image is generated, which destroys the continuity of image spatial information and increases the detection difficulty. The fused sample images are input into the pretrained text detection model and lightweight model, the distillation loss is calculated, and the parameters of the lightweight model are updated to improve the accuracy of text detection.

Benefits of technology

It improves the generalization of text detection models on difficult samples, enhances the efficiency of knowledge distillation, achieves higher text detection accuracy, and reduces data labeling costs and training cycles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119625704A_ABST
    Figure CN119625704A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of text detection, and discloses a training method of a text detection model, and a text detection method, device and equipment, and the method comprises the steps: obtaining a target sample image and a target image; fusing the target sample image and the target image to obtain a fused sample image; the fused sample image is input into a first text detection model and a second text detection model, a corresponding first text prediction result and a corresponding second text prediction result are obtained, and the first text detection model is obtained through pre-training; obtaining distillation loss based on the first text prediction result and the second text prediction result; and updating parameters of the second text detection model based on the distillation loss to determine a target text detection model. The problem of text detection accuracy can be solved.
Need to check novelty before this filing date? Find Prior Art