基于改进YOLOv5s的小目标检测算法
Small object detection algorithm based on improved YOLOv5s
贵向泉 1秦庆松 1孔令旺2
作者信息
- 1. 兰州理工大学计算机与通信学院,甘肃兰州 730050
- 2. 甘肃省气象信息与技术装备保障中心 气象数据质量控制室,甘肃兰州 730020
- 折叠
摘要
针对当前主流目标检测算法对图像中远距离小目标产生的漏检、误检等问题,提出一种改进YOLOv5s的小 目标检测算法.在模型训练过程中,通过引入Focal-EIOU定位损失函数,加强边界框的定位精度;在骨干网络中,通过添加小目标检测层,提高小目标的检测精度;在Neck结构中,通过优化上采样算子和添加注意力机制,加强小目标的特征信息.实验结果表明,改进后的算法在VisDrone数据集上与YOLOv5s算法相比,mAP@small提高了 3.2％,且检测速度满足实时性的要求,能够很好地应用于小目标检测任务中.
Abstract
Aiming at the problems of missed detection and false detection caused by the current mainstream object detection algo-rithms for small objects in the image at long distance,an improved YOLOv5s small object detection algorithm was proposed.In the process of model training,Focal-EIOU positioning loss function was introduced to enhance the positioning accuracy of the bounding box.In the backbone network,small object detection layer was added to improve the detection accuracy of small objects.In the Neck structure,the feature information of small objects was enhanced by optimizing the upper sampling operator and adding attention mechanism.Experimental results show that the improved algorithm is compared with YOLOv5s algorithm on VisDrone dataset,the mAP@small can be improved by 3.2％,and the detection speed meets the real-time requirements,which can be well applied to small object detection tasks.
关键词
YOLOv5s算法/小目标检测/损失函数/上采样算子/骨干网络/注意力机制/特征信息Key words
YOLOv5s algorithm/small object detection/loss function/upsampling operator/backbone network/attention mechanism/feature information引用本文复制引用
基金项目
甘肃省重点研发计划-工业类基金项目(22YF7GA159)
国家重点研发计划基金项目(2020YFB1713600)
出版年
2024
NETL
NSTL
万方数据