Automated Aircraft Surface Anomaly Identification Using Deep Learning-Based Object Detectors

Automated inspection of aircraft surfaces is a critical component of predictive maintenance strategies aimed at improving operational safety and reducing downtime in the aviation industry. In this study, we conduct a comprehensive evaluation of ten object detection algorithms, including Faster R-CNN, RetinaNet, SSD variants, RefineDet, CenterNet-HG, and several YOLO models, for the automated detection of surface defects such as cracks, dents, and corrosion. The evaluation is based on a high-resolution dataset of real-world aircraft imagery and encompasses both detection performance metrics and computational efficiency. Results show that YOLO-based models, particularly YOLOv5x and YOLOv8l, offer the best trade-off between accuracy and computational cost, significantly outperforming traditional two-stage and legacy one-stage detectors. Through clustering, scatter, and rank-based analyses, we highlight performance groupings and generalization behavior across varying IoU thresholds. This work demonstrates the effectiveness of modern deep learning models in automating aircraft defect detection and underscores their potential for deployment in intelligent maintenance systems.