A novel dual-student reverse knowledge distillation method for magnetic tile defect detection.

Area of Science:

• Materials Science
• Computer Vision
• Industrial Automation

Background:

• Surface defect detection in magnetic tiles is critical for industrial production.
• Challenges include similar defect/background textures, low contrast, and variable defect morphology.
• Existing methods struggle with these complexities.

Purpose of the Study:

• To propose a novel dual-student reverse knowledge distillation framework (BSDRD) for enhanced magnetic tile surface defect detection.
• To address challenges of similar textures, low contrast, and variable defect sizes/morphologies.
• To improve both pixel-level and sample-level anomaly detection accuracy.

Main Methods:

• Developed a Binary Struct and Detail Reverse Distillation (BSDRD) framework.
• Employed a pre-trained teacher network for multi-scale feature extraction.
• Utilized two student networks: one for feature fusion/compression, another for wavelet-based decomposition (struct and detail students).
• Introduced a multi-dimensional feature gated fusion loss (MD-GFLoss) for improved feature selectivity.

Main Results:

• The BSDRD framework demonstrated high effectiveness in detecting defects with complex textures and small sizes.
• Achieved superior performance compared to existing methods on magnetic tile defect detection datasets.
• Showcased significant improvements in both pixel-level and sample-level anomaly detection tasks.

Conclusions:

• The proposed BSDRD framework offers a robust solution for magnetic tile surface defect detection.
• The dual-student approach with wavelet decomposition enhances sensitivity to subtle and complex defects.
• BSDRD provides a promising advancement for quality control in industrial magnetic tile production.