Multi-Scale Deformable Transformer with Iterative Query Refinement for Hot-Rolled Steel Surface Defect Detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Detecting small, complex surface defects on hot-rolled steel plates is crucial for industrial quality assurance.
• Existing deep learning detectors struggle with accuracy and training speed for small object detection, hindering real-time applications.

Purpose of the Study:

• To propose a novel network, the multi-scale deformable transformer iterative query refinement network (MDT-Net), to address limitations in steel surface defect detection.
• To improve detection accuracy, computational efficiency, and training convergence speed for small and complex defects.

Main Methods:

• Developed MDT-Net, integrating a Swin Transformer backbone for multi-scale features.
• Incorporated a deformable attention mechanism to reduce computational complexity and accelerate convergence.
• Implemented an iterative bounding box refinement strategy for precise localization.

Main Results:

• MDT-Net achieved 82.7% mAP50 on the NEU-DET dataset.
• Demonstrated superior performance in small object detection, with an mAP50:95 of 0.55.
• Showcased significantly faster training convergence compared to mainstream detectors.

Conclusions:

• MDT-Net offers an effective and robust solution for accurate and efficient steel surface defect detection.
• The proposed network enhances sensor-based quality control and industrial quality management.
• MDT-Net provides a promising tool for real-time industrial inspection systems.