基于改进YOLOv5的航拍图像检测方法
An aerial image detection method based on improved YOLOv5
王嘉锵 1刘子德 1王绪娜 1高宏伟1
作者信息
- 1. 沈阳理工大学自动化与电气工程学院,辽宁沈阳 110159
- 折叠
摘要
由于航拍图像中的目标存在遮挡、重叠等问题,导致模型难以稳定识别,降低了军事目标追踪、交通监管、灾害观察等领域的工作效率.为解决上述问题,提出了一种基于改进YOLOv5的航拍图像检测方法.该方法引入用于低分辨率图像和小物体的新卷积神经网络模块(Space-to-depth Convolution,SPD-Conv)、小目标检测头、软非极大值抑制算法(Soft Non-maximum Suppression,Soft-NMS)和回归损失函数,并在VisDrone2019数据集上进行了大量实验.实验结果表明:所提方法在VisDrone2019数据集上平均准确率提高12.5%,mAP@0.5:0.95指标提高9.3%.
Abstract
Due to issues such as occlusion and overlapping in aerial images,it is challenging for models to achieve stable recogni-tion,which reduces efficiency in areas such as military target tracking,traffic monitoring,and disaster observation.To address these problems,a method based on improved YOLOv5 for aerial image detection has been proposed.This method introduces a new convolu-tional neural network module(Space-to-depth Convolution,SPD-Conv)for low-resolution images and small objects,a small object de-tection head,a soft non-maximum suppression algorithm(Soft Non-maximum Suppression,Soft-NMS),and a regression loss function.Extensive experiments have been conducted on the VisDrone2019 dataset.The experimental results show that the proposed method achieves an average accuracy improvement of 12.5%and a 9.3%increase in mAP@0.5:0.95 metric on the VisDrone2019 dataset.
关键词
小目标检测/SPD-Conv/Soft-NMS/回归损失函数Key words
Small target detection/SPD-Conv/Soft-NMS/Regression loss function引用本文复制引用
基金项目
辽宁省重点科技创新基地联合开放基金(2021-KF-12-05)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据