FDD: a deep learning-based steel defect detectors.

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Surface defects significantly impact industrial product quality.
• Automated inspection systems are crucial for addressing these defects efficiently.
• Existing methods often struggle with the geometric variability of defects.

Purpose of the Study:

• To introduce a novel deep learning-based system for steel surface defect detection.
• To enhance the accuracy and robustness of automated defect inspection.
• To improve overall manufacturing productivity and product quality.

Main Methods:

• Utilized a Cascade R-CNN baseline architecture.
• Incorporated deformable convolution and deformable RoI pooling for geometric adaptability.
• Implemented guided anchoring for precise bounding box generation.
• Introduced random and ultimate scaling techniques for enriched image perspectives.

Main Results:

• Demonstrated superior performance on Severstal, NEU, and DAGM steel defect datasets.
• Achieved significant improvements in average recall (AR) and mean average precision (mAP).
• Outperformed existing state-of-the-art defect detection methods.

Conclusions:

• The forceful defect detector (FDD) offers a highly effective solution for steel surface defect inspection.
• The proposed model contributes to accelerating industrial automation.
• The system sustains high product quality while increasing manufacturing productivity.