Enhanced RT-DETR With Multi-Scale Interaction and Hierarchical Attention for Aero-Engine Booster Vane Defect Detection

Automated defect detection for aero-engine booster vanes is critical yet challenging due to complex metallic textures and extreme scale variations. To address this, we propose an enhanced RT-DETR network integrating three novel modules: the Efficient Multi-Branch & Scale FPN (MBMS-FPN), Hierarchical Attention Fusion Block (HAFB), and Multi-scale Fusion Module (MFM). These components synergistically improve multi-scale feature representation and noise suppression. Experiments on our self-collected Aero-Blade-Defect (ABD) dataset demonstrate an mAP of 87.34%, representing an improvement of over 12% compared to the baseline and exceeding existing YOLO detectors. Furthermore, generalization is verified on the public NEU-DET benchmark with a competitive 79.15% mAP. Crucially, these results are obtained with a reasonable computational cost (60.6 GFLOPs), offering an effective efficiency-accuracy trade-off suitable for safety-critical aviation inspections.