A Co-Designed Framework Combining Dome-Aperture Imaging and Generative AI for Defect Detection on Non-Planar Metal...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Automated visual inspection of safety-critical metal assemblies is difficult due to complex shapes, reflective surfaces, and few defect examples.
• Traditional 2D methods face imaging issues and poor performance with limited data, while 3D methods are often slow and costly.

Purpose of the Study:

• To develop an integrated system for defect imaging, generation, and detection for challenging metal components.
• To enhance defect perception capabilities through co-designed imaging and deep generative models.

Main Methods:

• Developed a specialized imaging system with dome illumination and a small-aperture lens for non-planar metal surfaces.
• Implemented a dual-stage generative strategy using an improved FastGAN (DMGF-SLE) for defect patches and Poisson image editing for seamless fusion.
• Utilized perceptual and optimized loss functions to focus on realistic defect generation.

Main Results:

• Generated realistic defect samples efficiently under few-shot conditions, improving Fréchet Inception Distance (FID) scores by 11-24% over baseline models.
• Significantly enhanced downstream detection performance, increasing YOLOv8's mAP@50:95 from 50.4% to 60.5% using synthetic data.
• Demonstrated a computationally efficient and practically viable system solution.

Conclusions:

• The proposed integrated system offers a complete, synergistic, and deployable solution for defect detection in challenging metal components.
• This approach provides a viable technical pathway for improving automated visual inspection in manufacturing.
• The co-design of imaging and generative models effectively addresses limitations in current defect detection methods.