基于深度神经网络的多源遥感图像房屋检测技术研究

CASIA OpenIR > 毕业生 > 硕士学位论文

	基于深度神经网络的多源遥感图像房屋检测技术研究
	黄祖明
学位类型	工学硕士
导师	潘春洪
	2017
学位授予单位	中国科学院研究生院
学位授予地点	北京
关键词	多源遥感图像房屋检测深度神经网络公开地图项目
其他摘要	遥感图像中的房屋检测是目标检测领域的一个重要分支。其在城市规划、变化检测及 GIS 信息构建等方面发挥着至关重要的作用，日渐成为学术界和工业界的研究热点之一。然而，遥感图像的房屋具有许多复杂的特性，如多变的尺度、复杂的背景及多样的拓扑形状等。一般的目标检测算法难以直接应用于遥感图像的房屋检测任务中。因此，目前遥感图像中的房屋检测仍是一项亟需解决的挑战性课题。传统的房屋检测算法主要基于人工构造的特征，如形状特征、边缘特征及阴影特征等。这类算法仅能提取房屋目标的浅层特征，不能充分地表达其高层语义信息，准确率低，鲁棒性差。目前，由于深度神经网络被广泛地应用到目标检测任务中，如 AlexNet、 YOLO 及 FCN 等。房屋检测算法的主流趋势也逐渐倾向于使用深度神经网络。这类算法能利用高度的非线性映射及大数据训练，提取出房屋目标的深层特征，较好地克服了传统算法的缺陷。然而，由于深度神经网络在遥感图像的房屋检测任务中方兴未艾，许多基于深度神经网络的房屋检测算法在准确率、鲁棒性及实时性等方面还存在较大的提升空间。因此，如何结合深度神经网络和遥感图像中的房屋检测的特点来设计一个性能良好的房屋检测框架是一个需要深入研究的问题。其有着十分重要的理论研究价值和潜在的实际应用前景。为此，本文提出基于深度神经网络的多源遥感图像房屋检测算法。本文重点研究如何结合深度神经网络及遥感图像中的房屋检测的特点，设计一个准确率高、鲁棒性强、实时性好的房屋检测框架。本文的主要工作及贡献总结如下：提出了一种基于深度反卷积神经网络的多源遥感图像房屋检测算法。该算法主要由三部分组成。首先，将 2016 年 IEEE 数据融合竞赛提供的多源遥感图像数据进行预处理。基于公开地图项目，制作了一个高质量的温哥华房屋数据集。接着，分别在温哥华房屋数据集的两个谱段组合（RGB 和 NRG）上训练两个深度反卷积神经网络。最后，融合两个网络输出的得分图，得到最后的房屋检测结果。大量的实验表明，该算法能够准确并且有效地提取遥感图像中的房屋目标。提出了一种基于多任务神经网络级联的遥感图像房屋检测算法。该算法的一个优点是能够对遥感图像中的房屋目标进行实例区分。该算法的框架由三个部分组成，即实例区分网络、掩膜估计网络和多边形生成器。前两个网络组成一个级联的结构，共享卷积层的特征，能够进行端到端训练。多边形生成器根据掩膜估计网络的输出，生成表示房屋实例的多边形。在 2016 年SpaceNet 竞赛数据集上的大量实验表明，该算法能够取得良好的房屋检测性能。 ; Building detection from remote sensing images is of great importance in the field of object detection. It has drawn extensive attention both from Academy and Industry for its wide range of applications, such as urban planning, urban change detection and Geographic Information System (GIS) engineering. Nevertheless, building detection from remote sensing images has many complicated factors, such as various scales, complex background and rich topological appearances and so on. General object detection methods cannot be directly applied on this task. Thus, building detection from remote sensing images is still a quite challenging task. Traditional building detection algorithms are typically based on handcrafted features, such as SIFT, morphological profiles, color features, etc. These features can only extract the low-level representations of the buildings, which lead to inaccurate and vulnerable detection results. Currently, deep neural networks, such as AlexNet, YOLO, FCN, etc., are widely applied on the object detection task. The major trend of building detection algorithms is using deep neural networks. These algorithms can extract the high-level representations of the buildings via extra highly nonlinear mapping and large-scale data training, which could conquer the defects of the traditional building detection algorithms. However, since the applications of the deep neural networks on building detection remain relatively nascent, the accuracy, robustness and inference speed of the building detection frameworks still have much room for improvement. Therefore, how to combine deep neural networks with the domain knowledge of building detection from remote sensing images is a fundamental issue. Such case is worthy of being further studied to design a solid and effective building detection framework. That is the core motivation of our thesis. To address the issue above, our thesis propose novel algorithms for building detection from multi-source remote sensing images based on deep neural networks. Our research focus on how to combine deep neural networks with the domain knowledge of building detection from remote sensing images to design an accurate, robust and high inference speed building detection framework. Our works and contributions are summarized as follows: We propose a novel supervised building detection method via deep deconvolution neural networks (DeconvNet). Our method consists of three steps. First, we preprocess the multi-source remote sensing images provided by the 2016 IEEE GRSS Data Fusion Contest. A high-quality Vancouver building dataset is created on pan-sharpened images whose ground-truth are obtained from the OpenStreetMap project. Then, we train deep deconvolution networks on two band combinations (RGB and NRG) of our dataset, respectively. Moreover, the output score maps of the trained models are fused to produce the final building detection results. Extensive experiments on our Vancouver building dataset demonstrate the effectiveness and efficiency of the proposed method. A multi-task network cascades is proposed for building detection from remote sensing images. It should be noted that the proposed method is able to identify building instances. The algorithm consists of three parts, i.e., differentiating instances, estimating masks, and generating polygons. The first two networks, forming a cascaded structure, share their convolutional features and can be trained in an end-to-end fashion. The polygon generator takes the estimated masks as inputs, and generates polygon areas that represent buildings. Extensive experiments show high accuracy on the Rio de Janeiro building dataset provided by the 2016 SpaceNet Challenge.
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14679
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	黄祖明. 基于深度神经网络的多源遥感图像房屋检测技术研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
Huang_Building_detec（12700KB）	学位论文		暂不开放	CC BY-NC-SA	请求全文