Using ISU-GAN for unsupervised small sample defect detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Supervised deep learning methods for surface defect detection require extensive labeled data, which is costly and difficult to obtain in industrial settings.
• Existing methods struggle with the scarcity and high cost of labeling industrial defect samples.

Purpose of the Study:

• To propose a novel unsupervised small sample defect detection model (ISU-GAN) for industrial applications.
• To enhance the feature extraction capabilities of generative adversarial networks (GANs) for improved defect detection.
• To develop an accurate defect segmentation method compatible with GAN-based approaches.

Main Methods:

• Developed ISU-GAN based on the CycleGAN architecture, incorporating skip connections, SE modules, and Involution modules into the Generator.
• Implemented a Structural Similarity Index Measure (SSIM)-based defect segmentation method for precise contour extraction without post-processing.
• Utilized the DAGM2007 dataset for experimental validation.

Main Results:

• ISU-GAN achieved higher detection accuracy and finer defect profiles using less than 1/3 of the unlabeled training data compared to supervised models.
• Demonstrated improved detection accuracy (2.84% and 0.41%) and F1 scores (0.025 and 0.0012) over supervised models UNet and ResUNet++.
• The proposed method generated defect profiles closer to real profiles than comparative models.

Conclusions:

• Unsupervised learning with ISU-GAN offers a viable and effective solution for industrial surface defect detection with limited data.
• The enhanced Generator and SSIM-based segmentation significantly improve detection and segmentation accuracy.
• ISU-GAN provides a more accurate and efficient alternative to traditional supervised methods in resource-constrained industrial environments.