Aero-DETR: a frequency-adaptive sparse-interaction network for UAV-assisted aircraft surface defect detection

Abstract

Abstract Aircraft surface defect inspection is critical for aviation safety but remains challenging because defects in unmanned aerial vehicle (UAV) imagery are often small, low-contrast, and multi-scale. We construct the aircraft UAV defect inspection dataset (AUDID), containing 2,043 images and 5,669 bounding-box instances: 1,628 dents, 323 cracks, and 3,718 rivet abnormalities. Images from UAV field captures and airline maintenance records were annotated in LabelImg by trained annotators and cross-checked by experienced aircraft maintenance engineers. The training, validation, and test sets contain 1,633, 202, and 208 images, respectively. We further develop Aero-DETR (Aerial Detection Transformer), integrating frequency-adaptive feature enhancement, sparse interaction modeling, and multi-scale feature fusion. At a 640 $$\times$$ 640 resolution, Aero-DETR was compared with faster region-based convolutional neural network (Faster R-CNN), Single Shot MultiBox Detector (SSD), detection transformer (DETR), DETR with improved deNoising anchOr boxes (DINO), you only look once (YOLO), and real-time detection transformer (RT-DETR) variants. On aircraft defects version 2 and AUDID, Aero-DETR obtained mean average precision ( $$\textrm{mAP}_{50:95}$$ ) scores of 54.39% and 34.12%, outperforming RT-DETR with a ResNet-18 backbone (RT-DETR-r18) by 2.90 and 1.64 points, respectively, with 16.34% fewer giga floating-point operations (GFLOPs). On AUDID, it achieved the highest $$\textrm{mAP}_{50}$$ and $$\textrm{mAP}_{50:95}$$ and ran at 133.06 frames per second (FPS) on an NVIDIA A100.