ImbDef-GAN: Defect Image-Generation Method Based on Sample Imbalance.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Deep learning for defect detection requires extensive data, but defective samples are rare in industrial settings.
• Existing generative methods struggle with unnatural defect boundaries, mask misalignment, and incorrect defect placement.

Purpose of the Study:

• To introduce ImbDef-GAN, a novel generative framework addressing sample imbalance in defect detection.
• To improve the realism and diversity of generated defect images for training detection models.

Main Methods:

• A two-stage approach: background image generation (StyleGAN3 variant with Progress-coupled Gated Detail Injection) and conditioned defect image generation.
• Defect feature extraction with a residual branch and blending for natural transitions.
• Mask-aware discriminator, Edge Structure Loss, and Region Consistency Loss for fidelity.

Main Results:

• ImbDef-GAN generated more realistic and diverse defects compared to existing methods on the MVTec AD dataset.
• Training YOLOv11 with ImbDef-GAN generated data resulted in a 5.4% improvement in mAP@0.5.
• The framework effectively addresses unnatural transitions, mask misalignment, and out-of-bounds placement.

Conclusions:

• ImbDef-GAN offers a robust solution for generating synthetic defect data, mitigating challenges of sample scarcity in industrial defect detection.
• The proposed method significantly enhances the performance of downstream deep learning detectors, improving overall detection accuracy.