首页 > , Vol. , Issue () : 1993-2002
建筑物提取在城市规划等土地利用分析中发挥着重要作用。用于提取建筑物的传统方法通常基于手工特征和分类器，导致精度较低。本文基于卷积神经网络CNN(Convolutional Neural Networks),自主学习多级的和具有区分度的特征来更好地辨识建筑物和背景，实现航空影像中的像素级建筑物提取。此外，本文使用视野增强FoVE(Field-of-View Enhancement)方法减轻边缘现象(切片边缘附近的建筑物提取精度通常低于中心区域附近的精度)的影响，并进一步通过实验揭示FoVE的饱和性。通过在两个数据集上的实验表明，CNN能有效实现像素级建筑物提取，FoVE能进一步提高建筑物提取准确率。
Building extraction plays a significant role in land use analysis like urban planning. Conventional methods used for the extraction of buildings are usually based on the traditional hand-crafted features and the classical classifiers, which results in low accuracy. In this paper, a convolutional neural network (CNN) is designed to achieve the building extraction from aerial imagery in pixel level. The convolutional neural network is supposed to learn hierarchical and discriminative features to distinguish between buildings and backgrounds. Furthermore, a phenomenon that the prediction accuracy near the edges of a patch is usually lower than that near the central area is observed, which is called Marginal Phenomenon (MP) in this work. To alleviate the impact of the MP, the method of Field-of-View Enhancement (FoVE) is adopted. How the FoVE method impacts the building extraction is uncovered in more details through experiments. Experiments on two datasets reveal that the CNN can achieve a good performance on the building extraction and the FoVE can further improve the classification accuracy to some extent.