Intraclass Image Augmentation for Defect Detection Using Generative Adversarial Neural Networks.

Area of Science:

• Computer Vision
• Materials Science
• Artificial Intelligence

Background:

• Surface defect identification using computer vision faces challenges due to variations in lighting, noise, size, shape, and position.
• Intraclass variation in surface defects limits the generalization ability of current identification algorithms.

Purpose of the Study:

• To develop a pixel-level image augmentation method for improving surface defect identification.
• To enhance the generalization ability of computer vision models for defect detection.

Main Methods:

• A generative adversarial network (GAN) model, termed Magna-Defect-GAN, was developed for image-to-image translation.
• The model generates synthetic surface defect images conditioned on fine-grained labels, creating high intraclass diversity.
• The generated synthetic data was used to augment a magnetic particle inspection (MPI) dataset.

Main Results:

• Magna-Defect-GAN successfully generated realistic, high-resolution (512x512) synthetic surface defect images.
• The augmented dataset led to improved accuracy in the defect identification model.
• The proposed augmentation method demonstrated adaptability to other surface defect identification models.

Conclusions:

• The Magna-Defect-GAN approach effectively addresses the challenge of intraclass variation in surface defect identification.
• Pixel-level image augmentation using GANs is a viable strategy to boost the performance of defect detection systems.
• This method offers a controllable and adaptable solution for enhancing surface defect datasets.