DG-GAN: A High Quality Defect Image Generation Method for Defect Detection.

Area of Science:

• Materials Science
• Computer Vision
• Manufacturing Engineering

Background:

• Surface defect detection is critical for industrial product quality, safety, and efficiency.
• Insufficient defect image samples hinder the training of effective defect detection models.
• Existing methods struggle with data scarcity, impacting model performance.

Purpose of the Study:

• To propose a novel defect image generation method, DG-GAN, to address the challenge of limited defect sample data.
• To improve the training stability and generative capabilities of defect detection networks.
• To enhance the accuracy and convergence of defect detection models using generated data.

Main Methods:

• Developed DG-GAN, a progressive generative adversarial network for defect image synthesis.
• Incorporated D2 adversarial loss, cyclic consistency loss, a data augmentation module, and a self-attention mechanism.
• Validated the generated images' quality and diversity on two datasets and their impact on a YOLOX detection model.

Main Results:

• DG-GAN generated high-quality, diverse surface defect images, significantly reducing FID scores (mean reductions of 16.17 and 20.06).
• Training defect detection models with DG-GAN generated images improved convergence stability and detection accuracy.
• YOLOX detection accuracy saw significant increases (up to 6.1% and 20.4%) with added generated defect images.

Conclusions:

• DG-GAN effectively generates realistic surface defect images, overcoming data limitations in industrial settings.
• The generated images serve as valuable training data, boosting the performance of defect detection systems.
• DG-GAN demonstrates significant potential for enhancing industrial surface defect detection tasks.