基于YOLO的遥感影像目标快速检测轻量化网络研究  

Research on Lightweight Network for Rapid Detection of Remote Sensing Image Targets based on YOLO

作　　者：王伟[1] 程勇[2] 周玉科[3] 张文杰 王军[2] 何佳信 顾雅康 WANG Wei;CHENG Yong;ZHOU Yuke;ZHANG Wenjie;WANG Jun;HE Jiaxin;GU Yakang(School of Automation,Nanjing University of Information Science&Technology,Nanjing 210044,China;School of Software,Nanjing University of Information Science&Technology,Nanjing 210044,China;Key Laboratory of Ecosystem Network Observation and Modeling,Institute of Geographic and Nature Resources Research,Chinese Academy of Sciences,Beijing 100101,China;School of Geographical Sciences,Nanjing University of Information Science&Technology,Nanjing 210044,China;State Key Laboratory of Resources and Environmental Information System,Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]南京信息工程大学自动化学院,江苏南京210044 [2]南京信息工程大学软件学院,江苏南京210044 [3]中国科学院地理科学与资源研究所生态系统网络观测与模拟重点实验室,北京100101 [4]南京信息工程大学地理科学学院,江苏南京210044 [5]中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室,北京100101

出　　处：《遥感技术与应用》2024年第3期547-556,共10页Remote Sensing Technology and Application

基　　金：国家自然科学基金面上项目(41975183、41875184)。

摘　　要：基于高分辨率遥感影像的目标识别技术被广泛应用于国土资源监测和情报收集等领域,精确、快速的目标检测方法是目前遥感图像研究的热点与难点。然而,许多检测方法过于追求提升检测精度却忽略了检测速度。为此,基于YOLOX提出一种改进型轻量化网络,以实现检测速度和精度权衡。首先,针对特征提取主干模块,提出一种Mobilenetv3tiny网络,进行轻量化改进,通过减少网络参数量,提高检测速度;其次,在特征金字塔结构中引入Ghost网络,在保证检测精度的前提下,降低网络复杂性;最后,使用Alpha-IoU和VariFocal_Loss优化损失函数,提高网络收敛速度和定位精度。在NWPU VHR-10数据集进行消融实验,结果表明改进网络较基础网络,检测精度提升0.76%,速度提升19.72%,权重为11 M(Mega)且参数量减少65.66%,网络整体效果较好。此外,对DIOR数据集进行对比实验,在保证较高检测精度的同时,检测速度提高26.88%,证明了改进网络的有效性。因此,改进网络能够有效权衡检测速度和精度,易于设备部署,适用于遥感图像目标实时检测应用场景。Object recognition technology based on high-resolution remote sensing images is widely used in the fields of land and resource monitoring and intelligence collection.Accurate and fast object detection methods are the hot spots and difficulties in the current research on remote sensing images.However,the current detection methods overly pursue improving detection accuracy while ignoring detection speed.Therefore,an improved lightweight network is proposed based on YOLOX to balance detection speed and accuracy.Firstly,for the backbone of feature extraction,a Mobilenetv3tiny is proposed to improve the detection speed by reducing the pa⁃rameters of the network.Secondly,the Ghost is introduced into the feature pyramid networks to reduce the com⁃plexity of the network under the premise of ensuring detection accuracy.Finally,Alpha-IoU and VariFo⁃cal_Loss are used to optimize the loss function to improve the convergence speed and positioning accuracy of the network.The ablation experiment was carried out on the NWPU VHR-10 dataset.The results show that,com⁃pared with the baseline,the improved network has a detection accuracy increase of 0.76%,a speed increase of 19.72%,a weight of 11 M(Mega),and a parameter reduction of 65.66%.The overall effect of the improved network is better.In addition,comparative experiments on the DIOR dataset show that the detection speed is improved by 26.88%while ensuring high detection accuracy.And that proves the effectiveness of the improved network.Therefore,the improved network can effectively balance detection speed and accuracy and is easy to deploy,which makes it suitable for real-time detection of remote sensing image targets.

关 键 词：高分辨率遥感影像 目标检测 单阶段算法 轻量化网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]