| 摘 要: 针对远距离小尺度交通标志在下采样过程中细节易丢失的问题，本文提出一种非对称轻量化检测算法AFA-YOLO。该算法采用“浅层保真、深层控算”的层级非对称结构：在浅层高分辨率阶段引入Space-to-Depth操作以保留目标空间细节，在深层语义阶段采用步长为2的卷积控制通道扩展，并利用C3Ghost减少特征聚合过程中的参数冗余。同时，引入AS-CIoU损失函数改善小目标边界框回归性能。TT100K数据集实验表明，AFA-YOLO的参数量由2.59M降至2.38M，mAP50由69.4%提升至76.1%，统一ONNX Runtime测试下的前向推理速度达到334.51 FPS。尽管模型GFLOPs由6.6增加至9.6，但仍实现了检测精度、参数规模与推理效率之间的较好平衡。 |
| 关键词: 交通标志检测 小目标检测 轻量化网络 边缘计算 |
|
中图分类号:
文献标识码:
|
|
| An Asymmetric Lightweight Approach for Small-Scale Traffic Sign Detection |
|
wangxiao, zhuxiaohong
|
Wuhan University of Science and Technology
|
| Abstract: To address the loss of fine details of distant small-scale traffic signs during downsampling, this paper proposes an asymmetric lightweight detection algorithm, named AFA-YOLO. The proposed method adopts a hierarchical asymmetric design characterized by shallow-layer detail preservation and deep-layer computation control. Specifically, the Space-to-Depth operation is introduced at the shallow high-resolution stage to retain spatial details of small objects, while stride-2 convolution is employed in the deeper semantic stage to control channel expansion. In addition, C3Ghost is used to reduce parameter redundancy during feature aggregation. Furthermore, the AS-CIoU loss function is introduced to improve bounding box regression performance for small targets. Experiments on the TT100K dataset show that AFA-YOLO reduces the number of parameters from 2.59M to 2.38M and improves mAP50 from 69.4% to 76.1%. Under a unified ONNX Runtime inference setting, the proposed model achieves a forward inference speed of 334.51 FPS. Although the GFLOPs increase from 6.6 to 9.6, AFA-YOLO achieves a favorable balance among detection accuracy, model size, and practical inference efficiency. |
| Keywords: Traffic sign detection Small object detection Lightweight network Edge computing |
