Detection and classification of surface defects on hot-rolled steel using vision transformers

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Steel surface defects significantly impact material integrity and product quality.
• Automated defect detection is crucial for efficient quality control in steel manufacturing.
• Existing methods, like Convolutional Neural Networks (CNNs), have limitations in capturing complex defect patterns.

Purpose of the Study:

• To propose and evaluate a Vision Transformer (ViT) model for automated steel surface defect detection and classification.
• To identify the optimal hyperparameter configurations for the ViT model.
• To compare the performance of the ViT model against existing approaches.

Main Methods:

• Utilized an open-source image dataset of steel surfaces with six defect categories: crazing, inclusion, rolled in, pitted surface, scratches, and patches.
• Preprocessed defect images through resizing.
• Trained and evaluated a Vision Transformer model with various hyperparameter settings.
• Analyzed model performance using confusion matrices.

Main Results:

• The proposed Vision Transformer model achieved a high overall accuracy of 96.39% for defect detection and classification.
• Optimal hyperparameter configurations were identified, leading to superior classification performance.
• The ViT model demonstrated strong capabilities in distinguishing between different types of steel surface faults.

Conclusions:

• Vision Transformers are highly effective for real-time, complex steel surface defect detection and classification.
• ViTs offer a viable and potentially superior alternative to Convolutional Neural Networks (CNNs) for industrial quality monitoring.
• The developed model enables efficient and robust condition monitoring of steel surfaces.