重庆邮电大学学报（自然科学版）2025，Vol.37Issue(2)：185-195,11.DOI:10.3979/j.issn.1673-825X.202406120141
面向交通目标的多尺度轻量化检测模型
Multi-scale lightweight detection model for traffic targets
摘要
Abstract
To address the challenges of large object scale variations and densely packed targets in traffic object detection,this paper proposes a fast multiscale powerful-YOLO(FMP-YOLO)model.In the backbone network,a faster block mod-ule,designed with partial convolution(PConv)and fast Fourier convolution(FFC),is introduced to reduce redundant computations and memory access,enhance inference speed,and alleviate the issue of limited receptive fields.In the aggre-gation network,an improved group shuffle convolution(GSConv)replaces standard convolution to better capture multiscale features while further reducing model parameters and computational costs.Additionally,the conventional non-maximum suppression(NMS)algorithm is replaced with a combination of Powerful-IoU and soft non-maximum suppression(Soft-NMS),mitigating the feature learning degradation caused by reduced parameterization and improving model accuracy.Ex-perimental results on the SODA10M and MS COCO datasets demonstrate that the improved model outperforms the original YOLOv8s,achieving approximately a 40%reduction in parameters and computation while increasing mAP by 1.7%and 1.4%,respectively.FMP-YOLO surpasses other classic models in terms of both compactness and accuracy,making it highly practical for real-world applications.关键词
交通目标检测/部分卷积/轻量化/软非极大值抑制(SoftNMS)Key words
traffic target detection/partial convolution/lightweight/soft non-maximum suppression(SoftNMS)分类
电子信息工程引用本文复制引用
刘伯红,郝文瑞..面向交通目标的多尺度轻量化检测模型[J].重庆邮电大学学报（自然科学版）,2025,37(2):185-195,11.基金项目
国家自然科学基金项目(62272075) National Natural Science Foundation of China(62272075) （62272075）
