基于改进YOLOv7的卫星遥感影像多尺度目标检测方法  

Multi-Scale Object Detection in Satellite Images Based on Improved YOLOv7

作　　者：魏瑀皓 黄松 黄亚妮 WEI Yuhao;HUANG Song;HUANG Yani(Collage of Command and Control Engineering,Army Engineering University of PLA,Nanjing 210000,China;31678 Corps of PLA,Luzhou 646000,China;31305 Corps of PLA,Chengdu 610000,China)

机构地区：[1]陆军工程大学指挥控制工程学院,南京210000 [2]中国人民解放军31678部队,泸州646000 [3]中国人民解放军31305部队,成都610000

出　　处：《航天返回与遥感》2024年第2期153-162,共10页Spacecraft Recovery & Remote Sensing

基　　金：军队研究生资助课题(JY2022049)。

摘　　要：针对卫星遥感影像目标检测中,小目标检测精度低、漏检率高,以及实际应用场景中检测效率低等问题,文章提出一种基于改进YOLOv7(You Only Look Once)的卫星遥感影像多尺度目标检测方法。在检测网络中,聚焦提升小目标检测能力,添加类注意力机制的卷积模块(ConvNeXt Block,CNeB),提升对小目标细粒度特征的提取及利用能力;同时,提出后处理机制,通过建立小目标与大目标的相互关系,实现使用单个模型对多种尺度目标进行检测。实验结果表明,在TGRS-HRRSD数据集4个小目标上,改进后的检测模型相较原始YOLOv7在平均精确率均值指标上提升了16.6个百分点。在检测特定大目标任务中,通过后处理机制,在保持精度的条件下,相较YOLT(You Only Look Twice)时间减少了70%。相较于主流的面向遥感影像的检测方法,该方法在检测多尺度目标上,检测精度更高、速度更快。To address the problems of low detection accuracy,high missed detection rate in small target detection,and low detection efficiency in practical application scenarios in satellite remote sensing image target detection,a multi-scale target detection method based on improved YOLOv7 for satellite remote sensing images is proposed.In the detection network,the focus is on improving the detection capability of small targets by adding a ConvNeXt Block(CNeB)with class attention,which enhances the ability of extraction and utilization of finegrained features of small targets.At the same time,a post-processing mechanism is proposed to establish the mutual relationship between small and large targets,enabling the detection of multiple-scale targets using a single model.Experimental results show that on four small targets in the TGRS-HRRSD dataset,the improved detection model achieved an average improvement of 16.6%in mean average precision compared to the original YOLOv7.In specific large target detection tasks,the post-processing mechanism reduced the time by 70%compared to YOLT while maintaining accuracy.Compared to mainstream remote sensing image detection methods,this method is more accurate and faster in detecting multi-scale targets.

关 键 词：卫星遥感图像 卷积神经网络 目标检测 多尺度 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置] V19[自动化与计算机技术—控制科学与工程]