【作者】 胡敏；

【导师】 况润元；

【作者基本信息】 江西理工大学 ， 地图学与地理信息系统， 2020， 硕士

【摘要】 在高分辨率遥感影像中,由于城市地表建筑物所具有的高度复杂性,使得直接基于遥感影像进行建筑物提取一直是影像信息分析中的难点。由于面向对象的建筑物提取方法在实际应用中提取精度不高以及后处理工作量大等局限性,近年来基于深度学习算法对建筑物进行提取的技术在建筑物提取的各类算法中崭露头角。目前,大多数深度学习方法均采用全卷积网络(FCN)的变体,如U-Net、SegNet和RestNet等。这些网络具有显著改善模型性能和抽象特征提取能力的优势,从而能够准确地完成图像的目标分割与识别。为了充分利用高分影像中建筑物的全局和局部信息以更精确地对建筑物进行分割提取,本文提出了基于全卷积神经网络上针对边界约束的校正神经网络模型(Boundary Regulated Network,BR-Net),使得建筑物的边界提取更加清晰完整。该模型由共享后端和多任务预测模块组成,利用修改的U-Net和多任务框架,根据共享后端的一致特征生成分割图预测值和轮廓构建。通过对边界信息的规定,提高了模型性能。并将改进后的模型应用于国际开源WHU建筑物数据集上,通过与传统的U-Net、SegNet、RestNet模型进行对比实验。实验表明改进后的模型,对于图像中建筑物提取的平均精度达到92.19%以上。该精度相对于传统U-Net、SegNet、RestNet模型,分别提高37.28%、4.09%、1.86%。本文选用江西省九江区域的GF-2影像,通过人工标注、图像裁剪、分区域统计、顺向旋转、镜像翻转对原始数据增广等预处理,建立了包含了栅格标签形式和矢量边界形式的多格式建筑物数据集。本文将数据划分为由224×224和512×512两种规格的图片组成的样本集,并将该样本集细分为主城区、乡村区和混合区三类区域。通过对U-Net、SegNet、RestNet模型与BR-Net的提取结果进行精度对比,探寻了不同规格图片的不同区域的建筑物提取精度。实验结果表明:在224×224规格的图片集中,主城区区域的建筑物提取的平均精度可达84.17%以上。为进一步凸显边界校正网络模型对建筑物提取的准确性,本文突破了传统研究中仅基于RGB图像的局限,采用了:1)遥感影像的RGB波段组合图片集,2)NIRRG波段组合图片集,3)前两类三波段组合图片的混合图片集,来对图片中的建筑物进行提取,从而最大程度上利用了GF-2影像波段。实验结果表明:相对于传统的只采用三波段信息的建筑物提取结果,本文基于四波段信息的建筑物提取结果的平均精度在224×224和512×512两类规格图片上分别提升了1.87%和1.74%。本文提出的边界校正网络,能够有效提升城市建筑物信息提取的精度,为城市建筑物信息的规模化提取提供了有效的途径。更多还原

【Abstract】 In high-resolution remote sensing images,due to the high complexity of urban surface buildings,it is always difficult to extract buildings based on remote sensing images directly in image information analysis.Due to the limitations of object-oriented building extraction methods in practical applications,such as low extraction accuracy and large post-processing workload,the technology of extracting buildings based on deep learning algorithms has emerged in various building extraction algorithms in recent years.Currently,most deep learning methods use variants of fully convolutional networks(FCN),such as U-Net,Seg Net,and Rest Net.These networks have the advantages of significantly improving model performance and abstract feature extraction capabilities,so that they can accurately complete image segmentation and recognition.In order to make full use of the global and local information of buildings in high-resolution images to more accurately segment and extract buildings,this paper proposes a corrected neural network model based on boundary constraints on a fully convolutional neural network to make the boundary of the building extracted More clear and complete.The model is composed of a shared back-end and multi-task prediction module.It uses a modified U-Net and multi-task framework to generate segmentation graph prediction values and contours based on the shared features of the shared back-end.By specifying the boundary information,the model performance is improved.The improved model is applied to the international open source WHU building data set,and compared with the traditional U-Net,Seg Net,Rest Net models.Experiments show that the improved model has an average accuracy of more than 92.19% for building extraction in the image.Compared with traditional U-Net,Seg Net,and Rest Net models,the accuracy is improved by 37.28%,4.09%,and 1.86%,respectively.In this paper,the GF-2 image of Jiujiang area in Jiangxi Province is selected.Through manual preprocessing,image cropping,sub-regional statistics,forward rotation,and mirror flip to increase the original data and other preprocessing,a grid label form and a vector boundary form are established.Multi-format building dataset.In this paper,the data is divided into a sample set composed of 224×224 and 512×512 pictures,and the sample set is subdivided into three types of areas: main urban area,rural area and mixed area.By comparing the accuracy of the extraction results of the U-Net,Seg Net and Rest Net models with the improved network model proposed in this paper,the accuracy of building extraction in different areas with different specifications of pictures is explored.The experimental results show that the average accuracy of building extraction in the main urban area can reach more than 84.17% in the 224×224 image collection.In order to further highlight the accuracy of the boundary correction network model for building extraction,this paper breaks through the limitations of traditional research based on RGB images only,using: 1)RGB band combined picture sets of remote sensing images,2)NIRRG band combined picture sets,3)A mixed picture set of the first two types of three-band combined pictures to extract the buildings in the pictures,thereby making the most of the GF-2 image band.The experimental results show that the average accuracy of the building extraction results based on the four-band information is improved by 1.87% on the two types of specifications pictures based on the four-band information compared with the traditional building extraction results that only use three-band information And 1.74%.The boundary correction network proposed in this paper can effectively improve the accuracy of urban building information extraction and provide an effective way for the largescale extraction of urban building information.更多还原