Unsupervised Anomaly Detection on Metal Surfaces Based on Frequency Domain Information Fusion.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Industrial manufacturing demands high-quality metal products, necessitating robust surface defect detection.
• Current defect detection methods face limitations due to scarce defect samples, unpredictable characteristics, and metal grain interference.
• Accurate metal surface defect detection is crucial for ensuring product quality and manufacturing efficiency.

Purpose of the Study:

• To propose an unsupervised algorithm, FFnet, for enhanced metal surface defect detection.
• To introduce frequency domain features into unsupervised anomaly detection for improved accuracy.
• To address the challenges of limited defect data and complex surface textures.

Main Methods:

• Developed FFnet, an unsupervised algorithm integrating frequency domain information with spatial features.
• Implemented adaptive frequency domain feature enhancement to prioritize anomalies over textures.
• Utilized a scale-adaptive feature reconstruction module for effective spatial-frequency domain fusion.
• Incorporated a feature selection module to enhance anomaly detection and reduce computational load.

Main Results:

• FFnet demonstrated superior performance on a connecting rod surface image dataset compared to state-of-the-art methods.
• The algorithm achieved optimal results in generalization experiments on the Kolektor Surface-Defect Dataset 2.
• FFnet exhibited strong generalization capabilities across different datasets and defect types.

Conclusions:

• The proposed FFnet algorithm offers an effective unsupervised approach for metal surface defect detection.
• Fusion of spatial and frequency domain features significantly improves anomaly detection accuracy.
• FFnet shows promise for real-world industrial applications requiring reliable defect identification.