基于改进YOLOv5s的航拍目标检测算法  

Aerospace Object Detection Algorithm Based on Enhanced YOLOv5s Model

作　　者：卢逸霏 林凯鑫 邹文文 陈小兰 罗德林 蔡荣贵 LU Yifei;LIN Kaixin;ZOU Wenwen;CHEN Xiaolan;LUO Delin;TSAI Rong-Guei(New College of Engineering and Industry,Putian University,Putian 351100,China;College of Information Technology,Xiamen Huatian International Vocation Institute,Xiamen 361102,China;School of Art Design and Architecture,Zhangzhou College of Science and Technology,Zhangzhou 363202,China;College of Information Engineering,Quanzhou Ocean Institute,Quanzhou 362700,China)

机构地区：[1]莆田学院新工科产业学院,福建莆田351100 [2]厦门华天涉外职业技术学院信息技术学院,福建厦门361102 [3]漳州科技职业学院艺术设计与建筑学院,福建漳州363202 [4]泉州海洋职业学院信息工程学院,福建泉州362700

出　　处：《软件导刊》2025年第3期193-199,共7页Software Guide

基　　金：福建省自然科学基金项目(2020J01924);莆田市科技计划项目(2022GZ2001ptxy14)。

摘　　要：航拍图像中普遍存在目标尺寸微小、尺度多变以及背景繁杂等问题,从而导致YOLO系列算法的检测精度较低。为此,基于YOLOv5s提出YOLO-SC2算法。首先,在YOLOv5s的网络结构基础上添加细粒化卷积模块、融合基于Transformer架构的C3TR层,并采用C2F模块替换C3模块以增强目标特征信息提取能力;其次,通过替换Focal-EIoU损失函数,使用软非极大值抑制算法增强小目标检测准确性;最后,融合解耦合头,提高多目标下的检测精度。实验数据表明,在VisDrone2019公开数据集上进行实验,相较于原始模型,改进后的模型在P、mAP_0.5、mAP_0.5:0.95指标上分别实现了6.2%、8.2%和8.4%的提升。通过与其他算法对比,验证了该改进算法的有效性。The widespread issue in aerial imagery of small target sizes,variable scales,and complex backgrounds leads to low detection accuracy with the YOLO series algorithms.To address this,this paper proposes the YOLO-SC2 algorithm,which is based on the YOLOv5s architecture.The algorithm incorporates a fine-grained convolutional module and integrates a Transformer-based C3TR layer into the network structure.The C2F module replaces the C3 module to enhance the extraction of target feature information.Additionally,the Focal-EIoU loss function is replaced to improve the accuracy of small target detection,and the Soft-NMS algorithm is employed for this purpose.Finally,a decoupled head fusion is introduced to enhance detection accuracy under multiple target scenarios.Experimental results on the VisDrone2019 public dataset demonstrate that the proposed model achieves improvements of 6.2%,8.2%,and 8.4%in the P,mAP_0.5,and mAP_0.5:0.95 metrics,respectively,compared to the original model.Comparative analysis with other algorithms validates the effectiveness of the improved algorithm presented in this study.

关 键 词：YOLOv5s 细粒化卷积模块 C3TR 损失函数 软非极大值抑制算法 解耦合头 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]