VQGNet: An Unsupervised Defect Detection Approach for Complex Textured Steel Surfaces

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Defect detection on steel surfaces with complex textures is challenging due to limited defect samples and complex annotations.
• Accurate defect segmentation is difficult, especially with intricate surface patterns.

Purpose of the Study:

• To propose VQGNet, an unsupervised algorithm for simultaneous defect recognition and segmentation on steel surfaces.
• To address the challenges of limited data and complex texture in industrial defect detection.

Main Methods:

• Developed VQGNet, an unsupervised algorithm integrating aggregated attention and a classification-aided module.
• Employed multi-scale feature fusion and neighbor feature aggregation for confident anomaly map generation.
• Introduced an anomaly generation method for grayscale images to aid model learning.

Main Results:

• VQGNet achieved state-of-the-art performance on an industrial steel dataset.
• Achieved I-AUROC of 99.6%, I-F1 of 98.8%, P-AUROC of 97.0%, and P-F1 of 80.3%.
• Demonstrated robust generalization with ViT-Query on the Kolektor Surface-Defect Dataset.

Conclusions:

• VQGNet effectively segments and classifies defects by focusing on anomalous information over complex textures.
• The proposed methods enhance anomaly detection confidence and model learning capabilities.
• VQGNet offers a powerful solution for unsupervised defect detection and segmentation in industrial applications.