基于改进U-Net的高分辨率遥感图像目标提取  

Target Extraction of High-resolution Remote Sensing Images Based on Improved U-Net

作　　者：孙岩 吴熙曦 雷震 SUN Yan;WU Xixi;LEI Zhen(Army Academy of Armored Forces,Beijing 100072,China)

机构地区：[1]陆军装甲兵学院,北京100072

出　　处：《指挥与控制学报》2023年第5期596-605,共10页Journal of Command and Control

摘　　要：为支持对遥感图像中地物目标的快速识别,提出一种基于改进U-Net神经网络的目标提取算法,选用经典的深度学习神经网络U-Net作为主干网络,提出了一种改进的U-Net网络架构,在编码器部分添加密集连接减轻(wide-range attention unit,WRAU)的网络退化问题和添加宽范围注意单元更好地融合多尺度特征通道,并在Massachusetts以及DeepGlobe数据集上进行评估,实验结果验证了所提网络架构的性能,相较于U-Net、ResUNet、UNetPPL、E-Net、SegNet等网络的优势.探讨了深度学习在遥感图像目标检测领域未来的研究趋势.To support the fast recognition of ground targets in remote sensing images,an improved U-Net neural network-based target extraction algorithm is proposed,the classical deep learning neural network U-Net is selected as the backbone network,dense connections in the encoder part is added to alleviate the network degradation problem of Wide-Range Attention Unit(WRAU)and is added to better fuse multi-scale feature channels.Massachusetts and DeepGlobe datasets are evaluated.The experimental results validate that the superiority performance of WRAU-Net compared with that of U-Net,ResUNet,UNetPPL,E-Net,SegNet and other networks.Finally,the future research trends of deep learning in the field of remote sensing image target detection is discussed in the conclusion section.

关 键 词：目标检测 深度学习 卷积神经网络 遥感图像 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]