基于深度学习的高分辨率遥感影像建筑物提取

Building Extraction From High-Resolution Remote Sensing Images With Deep Learning

利用遥感影像提取建筑物具有广泛应用,例如城市规划、人口估计、灾害评估、军事侦查等。然而,由于高分辨率遥感影像中建筑物外观结构复杂、尺度多样,为建筑物提取任务带来巨大挑战。提出了一种基于密集特征金字塔融合网络的建筑物提取方法。该方法构建了编码器-解码器结构,在编码路径使用深度残差网络提取图像特征,并引入空洞卷积、金字塔池化模块等手段进一步增强特征提取能力,获取图像不同尺度上、下文信息。在横向连接过程中嵌入卷积块注意模块,有效区分特征之间的重要程度,突出有用特征,减少误判概率。在解码路径特征融合过程中,采用密集跳跃连接方式计算特征金字塔,有效聚合不同尺度、不同分辨率特征。在IAILD(Inria Aerial Image Labeling Dataset)建筑物检测数据集的实验结果表明,该方法在获取深层次语义信息的同时兼顾了图像细节信息,在重叠度和准确率两项指标上均有显著提升,在高分辨率遥感影像建筑物检测任务中取得了更好的实际效果,优于其他常用的语义分割网络。

Building extraction using remote sensing images has many practical applications,such as urban planning,population estimation,disaster management,and military reconnaissance.However,the complex appearance and various scales of buildings in high-resolution remote sensing images bring a challenge for building extraction.In this study,we developed a method of building extraction based on a dense feature pyramid fusion network.We built an encoder-decoder structure,and utilize deep residual network to extract the image features in the encoding path.The dilated convolutions and pyramid pooling module were introduced to improve the performance of feature extraction and obtain context information of images at various scales.Meanwhile,the convolutional block attention modules were embedded in lateral connection to distinguish the importance of different features,highlight the useful features,and reduce the probabilities of mis-classification.In the decoding path,the dense skip connection was used to calculate the feature pyramid and the features with different resolutions and scales were aggregated effectively.We performed experiments on IAILD building detection datasets.The results demonstrate that the method can obtain deep-layer semantic information and pay attention to detailed information.It also has a considerable improvement in intersection over union and accuracy.Additionally,the method has higher effect on building detection from high-resolution remote sensing images,and performs better than other common semantic segmentation networks.