YOLO-DyMiF:一种面向低算力平台的动态多尺度交通标志检测网络OA

To address the low detection accuracy caused by the small size of traffic signs in autonomous driving scenarios and their susceptibility to environmental interference,as well as the limited computing capability and power budget of onboard platforms that make complex models difficult to deploy,an improved lightweight detector named YOLO-DyMiF(Dynamic Mixer and Feature Fusion)is proposed.The proposed model,based on YOLOv10n,introduces two major improvements.Firstly,an Efficient Dynamic Mixer Structure(EDMS)based on Adaptive Efficient Convolution(AEConv)is designed and embedded into the C3k2 module to get a new module named C3k2_EDMS,which replaces the C2f module in YOLOv10n.This design effectively reduces the parameter scale while preserving the feature representation capability of the backbone network.Secondly,a dynamic feature fusion neck network is developed with the Hierarchical Multi-scale Spatial Enhancement(HMSE)module.Through cross-layer interactions and adaptive weighted fusion,the neck enhances multi-scale feature representation,improving the detection accuracy of small traffic signs while keeping detection performance on medium and large objects.Experimental results on the TT100K dataset show that,in comparison with the state-of-the-art Mamba-YOLOt,YOLO-DyMiF improves mAP50 by 1.0%,reduces the number of parameters by 58.3%,and decreases computational cost by 42.3%.The proposed model significantly reduces the computational cost while ensuring high detection accuracy,which provides reliable technical support for traffic sign detection in autonomous driving applications.