基于YOLOv5改进的航拍图像目标检测算法  

Improved Target Detection Algorithm for Aerial Images Based on YOLOv5

作　　者：郭业才[1,2] 孙京东 Amitave Saha Guo Yecai;Sun Jingdong;Amitave Saha(School of Electronic and Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Electronic Information Engineering,Wuxi University,Wuxi 214105,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,江苏南京210044 [2]无锡学院电子信息工程学院,江苏无锡214105

出　　处：《系统仿真学报》2025年第2期551-562,共12页Journal of System Simulation

基　　金：国家自然科学基金(61673222);江苏省研究生实践创新计划(SJCX23_0373)。

摘　　要：为解决小目标检测方法检测精度低、误检率高和漏检率高等问题,提出了FSD-YOLOv5算法,在YOLOv5算法的基础上进行了三方面的改进:用FocalEIoU代替了CIoU,提高了模型收敛速度和回归精度;针对CNN架构的缺陷,采用了一种新的CNN构建模块SPD-Conv;针对卷积神经网络降采样导致的特征图中小目标信息减少或丢失的问题,引入了特征重用来增加特征图中小目标的特征信息。仿真结果表明:FSD-YOLOv5的检测准确率为36.3%,比原算法提高了2.4%。In order to improve the existing small target detection methods,which suffer from low detection accuracy,high false detection rate and high leakage rate,the FSD-YOLOv5 algorithm is proposed,which has three improvements based on the YOLOv5 algorithm.The Focal EloU is used instead of the original CloU to improve the model convergence speed and regression accuracy.To cope with the deficiencies in CNN architecture,we adopt a new CNN building block called SPD-Conv is adopted.To address the problem of the reduced or lost information of small objects in feature maps caused by downsampling in convolutional neural networks,feature reuse is introduced to increase the feature information of small objects in the feature maps.Experimental results show that FSD-YOLOv5 achieves a detection accuracy of 36.3%,an improvement of 2.4%in comparison with original algorithm.

关 键 词：YOLOv5 FocalEIoU SPD-Conv 密集卷积网络 航拍图像检测 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]