A Ceramic Rare Defect Amplification Method Based on TC-CycleGAN.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning for ceramic defect detection faces challenges with limited rare defect samples and class imbalance.
• Existing generative image augmentation techniques struggle with poor image quality for rare defects, such as uneven brightness and insufficient small defect features.

Purpose of the Study:

• To propose an improved ceramic rare defect image augmentation method using TC-CycleGAN.
• To enhance the quality and diversity of generated ceramic defect images, particularly for rare defect types.

Main Methods:

• Developed TC-CycleGAN, optimizing the generator (TC-UNet with scSE and DehazeFormer) and discriminator (TC-PatchGAN with ContraNorm).
• Implemented novel modules within the generator and discriminator to better capture subtle and small-sized ceramic defect features.

Main Results:

• Significant improvements in generated image quality, with FID and KID score reductions of up to 73% for concave defects and 63% for smoke stains.
• Increased recognition accuracy for rare defects by 1.2% for concave types and 3.9% for smoke stains when using the augmented dataset for training.

Conclusions:

• The proposed TC-CycleGAN method effectively enhances the quality and diversity of augmented ceramic defect images.
• This augmentation technique substantially improves the performance of deep learning models in detecting rare ceramic defects.