A multi-scale pooling convolutional neural network for accurate steel surface defects classification

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Surface defect detection is crucial for industrial product quality inspection.
• Accurate classification of steel surface defects is essential for manufacturing processes.

Purpose of the Study:

• To develop an innovative multi-scale pooling convolutional neural network for high-accuracy steel surface defect classification.
• To evaluate the model's performance on both noise-free and noisy datasets.

Main Methods:

• A multi-scale pooling convolutional neural network, based on SqueezeNet, was developed.
• Experiments were conducted on the NEU noise-free and noisy testing sets.
• Class activation map and T-SNE visualization techniques were used for analysis.

Main Results:

• The multi-scale pooling model accurately captured defect locations at multiple scales, yielding robust results.
• T-SNE analysis showed large inter-class and small intra-class distances, indicating high reliability and generalization.
• The model is compact (3MB) and achieves high-speed inference (up to 130FPS on NVIDIA 1080Ti GPU).

Conclusions:

• The developed model offers a reliable and efficient solution for steel surface defect classification.
• Its real-time performance makes it suitable for industrial applications with high-speed requirements.
• The multi-scale approach enhances feature extraction for improved classification accuracy.