Anomaly detection improves UAV crack inspection for port quay walls
Researchers have improved UAV-based crack inspection for port quay walls by adding anomaly detection.
The approach targets a persistent barrier to automated port infrastructure maintenance. UAV image systems can find crack candidates quickly, but they often flag debris, stains and uneven surface textures as cracks, creating false positives that reduce confidence in field use.
The proposed framework uses aerial images captured by a small general-purpose UAV. It first applies object detection to identify candidate crack regions, then uses anomaly detection to reassess those regions based on how far their image features deviate from the learned feature distribution of crack images.
The anomaly detection model is based on a Vision Transformer, or ViT. Tests across five verification areas found that standard-threshold object detection combined with anomaly detection consistently improved F1 and F2 scores compared with the conventional baseline, showing stronger false-positive suppression while maintaining crack detectability.
The study also tested a low-threshold object detection setting. In some 5-meter cases, Frangi filter preprocessing outperformed grayscale preprocessing by balancing broader crack extraction with false-positive suppression, but the benefit declined as image resolution deteriorated. The findings indicate that the most robust current setup is standard-threshold object detection with anomaly-based false-positive reduction, while practical deployment will require threshold calibration for each site and ground sampling distance.