Knowledge distillation and pseudo-labeling for lightweight YOLOv11-based structural crack detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Manual inspection of structural cracks is inefficient and unreliable.
• Automated crack detection is crucial for infrastructure safety and durability.
• Lightweight deep learning models are needed for edge-device deployment.

Purpose of the Study:

• To improve a lightweight YOLOv11-N detector for crack detection using a training-centric strategy.
• To investigate semi-supervised learning mechanisms: pseudo-labeling and knowledge distillation.
• To validate the model's performance on edge devices for practical applications.

Main Methods:

• Knowledge transfer from a high-capacity YOLOv11-L teacher to a lightweight YOLOv11-N student.
• Semi-supervised learning: pseudo-labeling with IoU/NMS filtering and knowledge distillation.
• Systematic data augmentation to enlarge the crack dataset.

Main Results:

• Both pseudo-labeling and knowledge distillation enhanced the baseline student model.
• Pseudo-labeling demonstrated more stable training and stronger overall performance gains.
• Knowledge distillation improved convergence behavior and sample efficiency.
• Edge-device benchmarks confirmed suitability for resource-constrained platforms.

Conclusions:

• The proposed training-centric strategy effectively improves lightweight crack detection models.
• Pseudo-labeling offers robust performance gains, while distillation enhances training efficiency.
• The optimized lightweight detector is deployable on edge devices for real-time infrastructure monitoring.