Autoencoder-Based Unsupervised Surface Defect Detection Using Two-Stage Training

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Surface defect detection is crucial for quality control.
• Unsupervised methods face challenges in accurate defect identification and normal background reconstruction.
• Existing methods often struggle with noise and detail preservation.

Purpose of the Study:

• To develop an unsupervised method for accurate surface defect detection.
• To achieve high-quality normal background reconstruction without noise.
• To improve feature learning and detail preservation in defect detection.

Main Methods:

• Proposed an adaptive weighted structural similarity (AW-SSIM) loss function.
• Introduced an artificial defect generation algorithm (ADGA).
• Implemented a two-stage training strategy with combined AW-SSIM and learned Perceptual Image Patch Similarity (LPIPS) loss.

Main Results:

• Achieved accurate defect detection and high-quality normal background reconstruction.
• Demonstrated state-of-the-art defect detection accuracy.
• Obtained an average area under the receiver operating characteristic curve (AuROC) of 97.69% on the MVTec anomaly detection dataset.

Conclusions:

• The proposed unsupervised method effectively addresses challenges in surface defect detection.
• AW-SSIM loss and the two-stage training strategy significantly improve performance.
• The method shows strong generalization capabilities on diverse defect types.