改进YOLOv8的遥感图像飞机目标轻量化检测算法  

Improved YOLOv8 Remote Sensing Image Aircraft Target Lightweight Detection Algorithm

作　　者：王莹 张上 胡益民 王恒涛 院永莹 WANG Ying;ZHANG Shang;HU Yiming;WANG Hengtao;YUAN Yongying(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang,Hubei 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang,Hubei 443002,China)

机构地区：[1]水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《遥感信息》2025年第1期134-141,共8页Remote Sensing Information

摘　　要：针对遥感飞机目标检测存在模型体积和计算量较大的问题,提出一种基于YOLOv8的轻量化遥感飞机目标检测算法。首先,在主干网络中引入一种融合SRU和CRU的C2f_ScConv模块,通过空间重建和通道重建来减少中间特征映射冗余,并增强飞机目标特征表示,在保证精度的同时实现模型轻量化。然后,在头部网络中融入轻量型注意力机制shuffle attention,利用shuffle unit捕捉通道和空间维度的特征依赖,利用channel shuffle算子实现不同子特征之间的信息交流,进而提高飞机目标检测精度。最后,使用Inner-CIoU作为损失函数,加速样本收敛。实验表明,改进算法相较于YOLOv5s、YOLOv7-Tiny、YOLOv8n模型,体积降低57.2%、52.4%、7.9%,计算量降低53.4%、43.6%、7.4%,在保持较高模型检测精度的同时大幅降低了模型的复杂度,达到了轻量化的要求。A lightweight remote sensing aircraft target detection model based on YOLOv8 is proposed to solve the problems of large volume and calculation.Firstly,a C2f_ScConv module integrating SRU and CRU is introduced into the backbone network.This module reduces the redundancy of intermediate feature mapping through space reconstruction and channel reconstruction,and enhances the representation of aircraft target features,which ensures the accuracy and realizes the model lightweight.Then,shuffle attention,a lightweight attention mechanism,is integrated into the head network.Shuffle unit is used to capture feature dependencies of channels and spatial dimensions.Channel shuffle operator realizes information exchange between different sub-features.Finally,Inner-CIoU is adopted as a loss function to accelerate sample convergence.The experiment shows that:compared with YOLOv5s,YOLOv7-Tiny and YOLOv8n models,the improved algorithm reduces the volume by 57.2%,52.4%and 7.9%,and reduces the computation amount by 53.4%,43.6%and 7.4%,respectively,which greatly reduces the complexity of the model while maintaining high model detection accuracy,and meets the requirements of lightweight.

关 键 词：飞机目标 YOLOv8 轻量化 C2f_ScConv shuffle attention Inner-IoU 

分 类 号：V271.4[航空宇航科学与技术—飞行器设计] TP751[自动化与计算机技术—检测技术与自动化装置]