基于轻量化YOLO11-ALS模型的轨道扣件多模态图像检测OA

To meet high-precision and lightweight requirements of track fastener detection models in complex environments,an efficient and lightweight detection model,YOLO11-ALS,is proposed.It effectively detects three fastener states:normal,loose,and missing.The core innovation lies in the design of an asymptotic feature pyramid neck network(AFPNNet).Through an adaptive spatial weighting and asymptotic feature fusion,model complexity is significantly reduced while multi-scale feature representation is enhanced.An asymmetric decoupled detection head(LADH)is designed,which adopts a parallel-branch structure to decouple classification and localization tasks,achieving a balance between detection performance and efficiency.The SIoU loss function is employed,and multi-dimensional geometric constraints are introduced to improve bounding box localization accuracy and training stability.Additionally,a multimodal fastener image dataset,including color,grayscale,and 3D-rendered images,is constructed to improve model generalization.The experimental results show that,compared with the original YOLO11 baseline,the improved model achieves reductions of 14.5％,4.8％,and 15.7％in size,computational complexity,and parameter count respectively,while precision,recall,and F1 score are improved by 2.9％,1.12％,and 2.15％respectively,demonstrating a comprehensive improvement in both detection performance and model lightweighting.