A method and apparatus for clustering repetitive steel defect images

By constructing labeled and unlabeled defect sample sets, and combining generalized category discovery and hierarchical distribution alignment regularization, the accuracy and adaptability issues of repetitive defect clustering on steel surfaces are solved, achieving efficient identification and clustering in complex environments.

CN122135058APending Publication Date: 2026-06-02UNIV OF SCI & TECH BEIJING
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
UNIV OF SCI & TECH BEIJING
Filing Date
2026-02-05
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing methods for detecting defects on steel surfaces are difficult to effectively distinguish between repetitive defects with similar appearances in complex production environments, and the accuracy of defect clustering is low, especially when faced with new morphological defects.

Method used

A repetitive steel defect image clustering method is adopted. By acquiring steel surface image data, labeled and unlabeled defect sample sets are constructed. Generalized category discovery learning and hierarchical distribution alignment regularization are used to achieve stable identification of known defect clusters and automatic discovery of unknown defect clusters, thereby improving clustering accuracy.

Benefits of technology

In complex industrial environments, it improves the clustering accuracy and robustness of repetitive defects, can adapt to factors such as missed detection, false detection, and speed disturbance, reduces quality risks, and assists in tracing the source of equipment anomalies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122135058A_ABST
    Figure CN122135058A_ABST
Patent Text Reader

Abstract

This invention discloses a method and apparatus for repetitive steel defect image clustering, relating to the field of industrial vision technology. The method includes: acquiring steel surface image data collected during steel production, performing defect detection processing on the data, and extracting defect sample images containing defect regions; constructing a labeled defect sample set based on historical steel surface defect data and performing supervised training on a defect clustering model during the initialization phase to obtain an initialized defect clustering model; constructing an unlabeled defect sample set based on the steel surface image data collected during steel production; training the initialized defect clustering model in a category discovery phase based on the unlabeled and labeled defect sample sets to obtain a trained defect clustering model; inputting the defect sample images containing defect regions into the trained defect clustering model for cluster prediction, and outputting a defect cluster label corresponding to each defect sample. Using this invention can improve clustering accuracy.
Need to check novelty before this filing date? Find Prior Art