MobileSteelNet: A Lightweight Steel Surface Defect Classification Network with Cross-Interactive Efficient...

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Accurate steel surface defect classification is crucial for industrial quality control.
• Existing methods often fail to balance accuracy and efficiency for real-time vision-based systems.

Purpose of the Study:

• To develop a lightweight deep learning framework for efficient and accurate steel surface defect classification.
• To enable real-time deployment in vision sensor systems for steel manufacturing.

Main Methods:

• Proposed MobileSteelNet, a lightweight deep learning framework.
• Introduced novel Multi-Scale Feature Fusion (MSFF) and Cross-Interactive Efficient Multi-Scale Attention (CIEMA) modules.
• Implemented grouped efficient computation and parallel multi-scale spatial extraction.

Main Results:

• Achieved 91.36% average accuracy on the NEU-DET dataset, outperforming ResNet-50 and MobileNetV2.
• Reached 93.70% accuracy for Scratch-type defects, an 82.12 percentage point improvement over MobileNetV1.
• Model size is only 8.2 MB, suitable for edge deployment.

Conclusions:

• MobileSteelNet offers a superior balance of accuracy and efficiency for steel surface defect classification.
• The framework meets lightweight deployment requirements for edge vision sensor systems in steel manufacturing.
• Demonstrates significant improvements, particularly for challenging defect types like scratches.