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

Structural cracks threaten the safety and long-term durability of civil infrastructure, yet manual inspection remains slow, subjective, and unreliable on complex surfaces. This study presents a training-centric strategy to improve a lightweight YOLOv11-N detector (2.6 M parameters) by transferring knowledge from a high-capacity YOLOv11-L teacher, without modifying the student architecture. Two semi-supervised mechanisms are investigated: pseudo-labeling and knowledge distillation. Using a crack dataset enlarged via systematic data augmentation, pseudo-labeling combines high-confidence teacher predictions with available ground-truth annotations through IoU and NMS filtering, while the distillation approach guides the student using both hard labels and teacher-derived soft signals to strengthen localization behavior. Experimental results show that both strategies enhance the baseline student model, with pseudo-labeling providing more stable training dynamics and stronger overall gains, whereas distillation primarily improves convergence behavior and sample efficiency. Ablation analyses highlight that the benefit of pseudo-labeling is data-dependent and requires a minimum pseudo-label density to achieve consistent improvements. Finally, edge-device (Raspberry 3B + and Nvidia Jetson Nano Kit) benchmarks validate that the resulting lightweight detector is suitable for deployment on resource-constrained platforms, enabling practical UAV- and mobile-oriented crack inspection across diverse edge computing tiers.