Explainable hybrid AI CAD framework for advanced prediction of steel surface defects

Area of Science:

• Materials Science
• Artificial Intelligence
• Computer Vision

Background:

• Steel surface defect detection is crucial for industrial quality control.
• Traditional single-stage detectors struggle with defect localization and classification, especially for similar or irregular defects.

Purpose of the Study:

• To propose a novel, explainable, hybrid AI Computer-Aided Design (CAD) framework for steel surface defect detection.
• To overcome the limitations of single-stage detectors by separating detection and classification tasks.

Main Methods:

• A two-stage framework: detection using Fusion YOLO (integrating DCBS-YOLO, YOLOv9c, YOLOv8s) for class-agnostic localization.
• Classification using a hybrid CNN-Vision Transformer (ViT) model for local texture and global dependency analysis.
• MLOps-based auto hyperparameter tuning and Grad-CAM for explainability.

Main Results:

• Fusion YOLO achieved an 83.8% Average Precision (AP) on the NEU-DET dataset.
• The classification stage reached a 99.7% F1-score on the NEU-DET dataset.
• Validated generalization on GC10-DET: 71.5% mAP (detection) and 94.8% F1-score (classification).

Conclusions:

• The proposed hybrid AI framework significantly improves steel surface defect detection accuracy and robustness.
• The two-stage approach effectively addresses the trade-off between localization and classification.
• The explainable framework offers reliable industrial inspection capabilities.