Training method and device of image processing model

By decoupling the feature representations of classification and localization tasks, and employing an unsupervised domain-adaptive target detection algorithm, the problem of data distribution differences during the training and deployment phases of image processing models is solved, thereby improving the model's generalization ability and robustness.

CN115205611BActive Publication Date: 2026-06-16BEIJING WODONG TIANJUN INFORMATION TECH CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
BEIJING WODONG TIANJUN INFORMATION TECH CO LTD
Filing Date
2021-04-12
Publication Date
2026-06-16

AI Technical Summary

⚠Technical Problem

In existing technologies, the difference in data distribution between the training and deployment phases of image processing models leads to performance degradation and makes them unable to adapt to changes in data distribution in real-world application scenarios.

⚗Method used

An unsupervised domain-adaptive target detection algorithm is adopted. By decoupling the feature representations of classification and localization tasks, different model branches are used to learn domain-invariant features and conduct adversarial training to reduce the domain margin and improve the performance of the model in the deployment phase.

🎯Benefits of technology

It effectively eliminates the competitive influence between classification and localization tasks, improves the model's generalization ability and robustness under different real-world conditions, and enhances the model's performance during the deployment phase.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115205611B_ABST
    Figure CN115205611B_ABST
Patent Text Reader

Abstract

The present disclosure relates to a training method and device of an image processing model, and relates to the technical field of artificial intelligence. The training method comprises: using a first domain discriminator to distinguish whether image classification features extracted by a classification feature extractor belong to source domain images or target domain images; determining a classification loss function according to a discrimination result of the first domain discriminator and a domain label result of each classification feature; using a localizer to output target localization results of each source domain image according to localization features of each source domain image extracted by a localization feature extractor, and using a classifier to output target classification results of each source domain image according to image classification features of each source domain image; determining a source domain loss function according to the target classification results, the target localization results and training labels of each source domain image; and training a machine learning model according to the classification loss function and the source domain loss function.
Need to check novelty before this filing date? Find Prior Art