Category-Agnostic Cluster-Guided Selective Suppression for Weakly Supervised Surface Defect Localization

Shuai Jiang       Min Liu†      Yuxi Liu      Yunfeng Ma      Yaonan Wang     
Hunan University
† Corresponding Author
IEEE TIM 2025
Paper Code
pipeline
Visual comparisons of localization maps produced by vanilla CAM [1] and CSS for different samples (a)–(e)

CSS enable weakly supervised defect localization by Cluster-Guided Selective Suppression, characterized by:

  • Re-establishment of evaluation criteria for defect localization performance
  • Clustering algorithm is introduced for unsupervised class activation mapping
  • Intermittent two-stage training strategy

For application, a real-world surface defect inspection platform is developed.

Inspection Platform

pipeline
Diagram and simulation of the surface defect inspection platform, including image acquisition through full coverage scanning and defect inspection of aero-engine blades

Abstract

Surface defect localization is indispensable for the quality control of industrial products in the manufacturing process, supervised dense prediction of defect areas requires laborious and difficult pixel-level annotations. On the contrary, weakly supervised defect localization (WSDL) is much more practicable, and the class activation maps (CAMs) generated with image-level labels can serve as prior knowledge or cues for downstream tasks. However, due to interference of unforeseen background objects that may have high confidence in activation maps instead, the localization maps obtained from general WSDL methods are prone to be unreliable, which leads to errors or even missing localization. Therefore, in this article, we innovatively propose a simple, yet effective cluster-guided selective suppression (CSS) strategy for weakly supervised surface defect localization, aiming at the rectification of CAMs, thus producing more reliable localization maps. Specifically, CSS leverages converted rectangle patch pairs generated from CAMs and performs unsupervised vector-level cluster analysis, all patches are classified into defective or nondefective clusters. Subsequently, a categoryagnostic cluster is selected to provide suppression feedback for the image suppressor, and the suppressed images are recycled for feature re-learning. Extensive experimental results demonstrate that our proposed method achieves state-of-theart WSDL performance on three public datasets of metal and magnetic-tile surfaces, with average classification and localization accuracy improved by 7.19% and 5.40%, respectively.

Framework

All high-confidence regions are combined with their mapped original image patches to form a pair, the patch clustering process classifies patch pairs into a fixed number of clusters, and the patches in the selected category-agnostic cluster will be suppressed and fed into the backbone network for feature re-learning. At the inference time, only the trained backbone network is leveraged for defect localization.

pipeline

Results

Defect Classification

pipeline

Defect Localization

pipeline

Visualization

pipeline

Citation

@article{jiang2025category,
  title={Category-agnostic Cluster-guided Selective Suppression for Weakly Supervised Surface Defect Localization},
  author={Jiang, Shuai and Liu, Min and Liu, Yuxi and Ma, Yunfeng and Wang, Yaonan},
  journal={IEEE Transactions on Instrumentation and Measurement},
  year={2025},
  publisher={IEEE}
}