Knowledge-Distilled YOLOv11 for Structural Crack: Lightweight Surface Flaw with Pseudo-Label Guidance Detection

Structural cracks reduce the safety and durability of buildings and civil infrastructure, making early detection an important engineering need. Manual inspection is slow, subjective, and often fails to identify fine cracks under complex surface conditions. In this study, a lightweight YOLOv11-N model was improved without modifying its architecture by transferring knowledge from a high-capacity YOLOv11-L model. Two different strategies were applied: Pseudo-Labeling and Knowledge Distillation. The dataset, obtained from Roboflow, contains 2,121 images of cracks on concrete, asphalt and painted surfaces. In the first method, the teacher model generated high-confidence pseudo-labels, which were combined with ground-truth through IoU and NMS filters to train the student. In the second method, the student learned from both real labels and soft feature outputs of the teacher to imitate its localization behavior. Results showed that Pseudo-Labeling produced extremely stable training curves and performance metrics almost identical to the teacher model. Knowledge Distillation improved convergence speed and sample efficiency, although it exhibited metric fluctuations. The proposed approach demonstrates that high crack detection accuracy can be achieved with low hardware cost, enabling deployment on drones and mobile devices. Future work will focus on combining both strategies and testing real-time performance on edge systems.