深度学习无人机遥感影像房屋检测方法研究  

Research on Deep Learning UAV Remote Sensing Image House Detection Method

作　　者：王静[1] WANG Jing(Shandong Institute of Land Surveying and Mapping,Ji′nan 250013,China)

机构地区：[1]山东省国土测绘院,山东济南250013

出　　处：《测绘与空间地理信息》2023年第9期139-142,145,共5页Geomatics & Spatial Information Technology

摘　　要：针对无人机遥感影像中背景复杂、房屋目标尺度差异较大等特点,提出一种基于卷积神经网络的房屋目标检测方法。首先,使用全卷积神经网络构建特征提取骨干网络来解决模型在特征提取过程中的语义信息丢失问题,同时在特征提取层间采用添加批量重整化层来加快模型收敛,然后,通过一个4层的多尺度特征提取层来获取语义信息增强后的多个尺度特征图来实现对不同大小目标的检测输出。通过本文构建的数据预处理模块对训练集合进行增强处理后,对本文所提出方法以及Yolo-v4,R-FCN几种模型进行训练。实验结果表明,本文所提出方法检测精度可以达到85.13%,同时检测速度也可以达到32 fps,并且对于多种场景下的目标均具备良好的泛化能力。Aiming at the complex background of UAV remote sensing images and the large difference in the scale of house targets,a house target detection method based on convolutional neural network is proposed.First,a full convolutional neural network is used to construct a feature extraction backbone network to reduce the loss of semantic information of the model in the feature extraction process.At the same time,a batch renormalization layer is added between the feature extraction layers to speed up the model conver-gence,and then a four-layer the multi-scale feature extraction layer obtains multiple-scale feature maps with enhanced semantic in-formation to achieve detection output of targets of different sizes.After the training set is enhanced by the data preprocessing module constructed in this article,the method proposed in this article and the Yolo-v4 and R-FCN models are trained.Experimental results show that the detection accuracy of the proposed method can reach 85.13%,and the detection speed can also reach 32 fps,and it has good generalization ability for targets in a variety of scenarios.

关 键 词：卷积神经网络 目标检测 无人机 遥感影像 图像增强 

分 类 号：P237[天文地球—摄影测量与遥感]