联合膨胀卷积残差网络和金字塔池化表达的高分影像建筑物自动识别  

Building Extraction from High Resolution Remote Sensing Images by Combining Dilated Convolutional Residual Networks and Pyramid Pooling Representation

作　　者：乔文凡 慎利[1,2] 戴延帅 曹云刚 QIAO Wen-fan;SHEN Li;DAI Yan-shuai;CAO Yun-gang(State-Province Joint Engineering Laboratory of Spatial Information Technology for High-Speed Railway Safety,Southwest Jiaotong University,Chengdu 611756;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学高速铁路运营安全空间信息技术国家地方联合工程实验室,四川成都611756 [2]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《地理与地理信息科学》2018年第5期56-62,共7页Geography and Geo-Information Science

基　　金：国家重点研发计划项目(2016YFB0501403);国家自然科学基金项目(41401374);国家重点基础研究发展计划项目(2012CB719901);中央高校基本科研业务费专项(2682016CX079);国土资源部航空地球物理与遥感地质重点实验室开放基金项目(2016YLF10)

摘　　要：针对传统建筑物提取方法对视觉特征人为设计的依赖,以及基于全卷积神经网络模型对提取目标边缘轮廓保真度差和对不同粒度建筑物自适应提取弱等问题,该文提出一种联合膨胀卷积残差网络和金字塔池化表达的高分辨率遥感影像建筑物自动识别方法,其所构建的全卷积神经网络包括膨胀卷积残差网络和金字塔池化单元两部分。在残差网络中,通过膨胀卷积限制模型中特征图分辨率的严重损失,从而有效地保留更多的细节特征;在金字塔池化单元中,通过全局平均池化将特征图池化为不同尺度,并与原始的输入特征图相融合,形成多尺度特征表达。基于马萨诸塞州地区具有复杂地表覆盖的公开遥感影像数据集开展的实验表明,相比目前较为流行的几种全卷积神经网络分类方法,该文所提出的联合膨胀卷积残差网络和金字塔池化表达方法的提取精度更高,建筑物提取结果能够有效地保留边界的细节轮廓信息,同时对不同形状大小建筑物的自适应提取能力更强。Traditional building extraction approaches using high resolution remote sensing images rely on the handcrafted features,and the common methods based on fully convolutional neural networks often fail to capture the objects′boundary to delineate extraction details and also cannot achieve self-adaptive extraction for different buildings with various scales.To address this problem,an approach for automatic building extraction from high resolution remote sensing images is proposed by combining dilated convolutional residual network and pyramid pooling representation.The proposed approach builds a fully convolutional network,which consists of two components,i.e.,a dilated residual network and a pyramid pooling unit.In the dilated residual network,dilated convolution is adopted to avoid reducing the resolution of feature maps,thus maintaining more object boundary information to delineate segmentation details.In the pyramid pooling unit,global average pooling is used to pool the feature maps outputted by dilated residual network to feature maps of different scales,which are then combined with the original feature maps to generate multi-scale representation of visual features.Experiments results using the public Massachusetts building dataset,which include a number of high resolution remote sensing images with complex landscapes,indicate that the proposed approach outperforms three state-of-the-art methods in the aspect of the extraction accuracy of building.Furthermore,the extraction results of the proposed approach can preserve details along the building boundary better,and achieve self-adaptive extraction for different buildings with various sizes and shapes.

关 键 词：高分辨率遥感 建筑物识别 全卷积神经网络 金字塔池化 多尺度表达 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]