LSN-YOLO: A Defect Detection Method for Aeroengine Blades

To address the challenges of class imbalance and low detection accuracy for small defects in aero-engine blades surface inspection, this study proposes LSN-YOLO, an improved YOLOv10 network incorporating three key innovations: a Linear Deformable Convolution (LDConv) module for flexible feature extraction, a parameter-free SimAM attention mechanism and a novel NWD-Loss function that better characterizes pixel-wise weight distributions within bounding boxes. Experimental results on the AeBAD dataset demonstrate LSN-YOLO's superior performance, achieving 72.4 % precision (5.9 % improvement over baseline) and 70.4 % mAP50 (4.7 % improvement) while maintaining competitive inference speeds and lightweight architecture. These advancements establish LSN-YOLO as an efficient and practical solution for industrial aero-engine blades defect detection, effectively balancing accuracy and processing speed requirements.