Metal Surface Defect Detection Based on a Transformer with Multi-Scale Mask Feature Fusion

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Surface defect detection is critical for metal industrial product quality.
• Acquiring sufficient defective samples for supervised learning is challenging.
• Semi-supervised learning offers a viable solution for defect detection.

Purpose of the Study:

• To propose a novel semi-supervised method for metal surface defect detection.
• To address the challenge of limited defective sample availability.
• To enhance the accuracy and efficiency of detecting defects such as rust and scratches.

Main Methods:

• A Vision Transformer (ViT) integrated into a generative adversarial network (GAN).
• Multi-scale masked feature fusion using block masks during training and testing.
• Incorporation of token merging (ToMe) for improved training speed.
• An encoder-decoder network with long skip connections for feature fusion.

Main Results:

• The proposed method demonstrates superior performance on five metal industrial product datasets and the MVTec AD dataset.
• Effective learning of semantic context from normal samples for defect identification.
• Successful detection of various metal surface defects including rust and scratches.

Conclusions:

• The developed semi-supervised Transformer-based method significantly improves metal surface defect detection.
• The approach is robust and efficient, overcoming limitations of traditional methods and data scarcity.
• This technique offers a promising solution for quality control in metal industrial production.