改进YOLOv8的交通标志检测模型  

Improved YOLOv8 model for traffic sign detection

作　　者：赵艳芹 赵文栋 Zhao Yanqin;Zhao Wendong(School of Computer&Information Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学计算机与信息工程学院,哈尔滨150022

出　　处：《黑龙江科技大学学报》2025年第2期344-348,共5页Journal of Heilongjiang University of Science And Technology

基　　金：黑龙江省省属高等学校基本科研业务费项目(2022-KYYWF-0565)。

摘　　要：为解决交通标志在复杂背景下检测精度低,小目标交通标志的漏检和误检,模型体积过大难以部署等问题,建立一种改进的YOLOv8交通标志检测模型。设计轻量级C2fRVB模块替换原模型的C2f模块,结合RepViTBlock增强全局特征提取能力,降低参数量,引入小目标检测层,融合浅层特征细节以提升小目标识别效果。采用ADown下采样替代传统卷积下采样,减少特征图信息的损失,提高检测精度。结果表明,改进后的模型在数据集TT100K中的检测精度、召回率和平均精度分别达到78.8%、69.2%和77.1%,相比YOLOv8n模型提升了7.9%、5.5%和7.2%,参数量下降38.9%,模型在实现轻量化的同时增加了识别准确性。This paper is intended to address the low detection accuracy of the traffic signs in complex backgrounds,the missed and fault detections of small-target traffic signs,and the hard deployment due to oversized model,and proposes an improved traffic sign detection model based on YOLOv8.The study consists of designing a lightweight C2fRVB module to replace the original C2f module,enhancing the global feature extraction capabilities,reducing the parameters by RepViTBlock,introducing a small-target detection layer,and integrating shallow feature details to improve recognition of small targets;and adopting ADown downsampling instead of traditional convolutional downsampling to minimize the loss of feature map information,and boost the detection accuracy.The results demonstrate that the improved model enables the precision,recall,and average accuracy of 78.8%,69.2%,and 77.1%respectively in the TT100K dataset,representing the improvements of 7.9%,5.5%,and 7.2%over against the YOLOv8n model,while reducing parameters by 38.9%.The optimized model achieves both lightweight design and better recognition accuracy.

关 键 词：YOLOv8 交通标志 RepViT 小目标检测层 下采样 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]