基于单阶段神经网络的房屋屋顶检测方法  

The Roof Detection Method Based on Single Stage Neural Network

作　　者：彭劲松 PENG Jingsong(College of Landscape Architecture, Hunan Vocational College of Environmental Biology, Hengyang Hunan 421005, China)

机构地区：[1]湖南环境生物职业技术学院园林学院,湖南衡阳421005

出　　处：《北京测绘》2021年第8期1089-1092,共4页Beijing Surveying and Mapping

基　　金：湖南省自然基金(2019JJ6028)。

摘　　要：针对现有方法利用无人机影像对房屋目标进行检测的过程中存在错检、漏检率高等问题,构建了单阶段卷积神经网络来实施无人机影像房屋检测,并在德国宇航中心开源数据集(Deutsches Zentrum für Luft-und Raumfahrt,DLR)3K Vehicle的基础上,采用多种数字图像增强手段对原始图像进行数据增强处理,提高训练后模型的泛化能力。在测试数据集上对训练后的网络进行测试,采用精度均值(Average Precision,AP)和每秒传输帧数(Frames Per Second,FPS)指标进行评价。并将检测结果与经典的目标检测模型单激发多盒探测器(Single Shot Multi-Box Detector,SSD)以及YOLOv3的检测结果进行对比。结果表明,所构建的卷积神经网络对于无人机影像中的房屋目标尤其是小目标有着较高的检测精度,检测精度可以达到91.3%AP,相比SSD和YOLOv3在精度方面提高了11.5%和8.3%。同时网络的检测速度可以达到每秒传输帧数21 m·s^(-1),能够快速精确地检测出无人机影像中的房屋目标。In order to solve the problems of false detection and high rate of missed detection in the process of using UAV image to detect the house target,a single stage convolution neural network was constructed to implement the house detection of UAV image.On the basis of DLR 3K Vehicle,a variety of digital image enhancement methods were applied to enhance the original image data to improve the generalization ability of the training model.The trained network was tested on the data set and evaluated by Average Precision(AP)and Frames Per Second(FPS).The detection results were compared with those of Single Shot Multi-Box Detector(SSD)and YOLOv3.The results showed that the constructed convolutional neural network had a high detection accuracy for house targets,especially small targets in UAV images.The detection accuracy could reach 91.3%AP,which was 11.5%and 8.3%higher than SSD and YOLOv3.At the same time,the detection speed reached 21 m·s^(-1),which could quickly and accurately detect the house target in the UAV image.

关 键 词：深度学习 无人机影像 房屋检测 卷积神经网络 数据增强 

分 类 号：P231[天文地球—摄影测量与遥感]