Mean shift based prototypical network for steel surface anomaly recognition.

Area of Science:

• Materials Science and Engineering
• Computer Vision and Machine Learning

Background:

• Surface defects in hot-rolled steel sheets critically impact product quality and market acceptance.
• Early detection of these defects is essential for efficient steel production.

Purpose of the Study:

• To introduce and evaluate the Mean Shift based Prototypical network (MSPro-Net) for enhanced steel surface defect classification.
• To demonstrate MSPro-Net's superiority over classical Prototypical networks, particularly in few-shot learning contexts.

Main Methods:

• Development of MSPro-Net utilizing adaptive prototype computation via the mean-shift method.
• Comparative analysis against the classical Prototypical network (CL-ProNet) using multiple N-way K-shot scenarios.
• Validation on benchmark datasets including NEU, XSDD, and GC10-Det.

Main Results:

• MSPro-Net achieved significantly higher accuracy in defect classification compared to CL-ProNet across various datasets and few-shot settings.
• On the NEU dataset (6-way 25-shot), MSPro-Net reached 98.67% accuracy versus 50.67% for CL-ProNet.
• Superior performance was also observed on XSDD (96.00%) and GC10-Det (90.00%) datasets, demonstrating robustness.

Conclusions:

• MSPro-Net offers a more representative class prototype computation, leading to substantial improvements in steel surface defect detection.
• The proposed method excels in few-shot learning, providing a powerful tool for early defect identification in manufacturing.