Abstract: As remote sensing images cover wide ranges and contain complex scenes, the conventional building extraction methods based on convolution neural networks which have few layers, cannot fully excavate the abstract information of images. It results in low accuracy and high false detection rate. However, simply increasing the number of layers in the network will lead to the vanishing of the gradient flow and the dispersion of information flow. A new densely connected deep deconvolutional neural network is proposed in this paper, which introduces dense connections to deconvolution networks. This network which has 51 layers can automatically learn hierarchical features of images and can be trained end to end in order to avoid the vanishing of information flow in deep network. At the same time, with the deconvolution, it realizes building extraction in pixel level and has a good performance in extracting buildings. It works well in practice on the ISPRS 2D Semantic Labeling Contest data.

Key words: building extraction, remote sensing image, dense connection, convolutional neural network, deconvolution

摘要： 遥感图像覆盖范围广、场景复杂，目前基于卷积神经网络的建筑物提取方法因层数较少，不能充分挖掘图像的抽象信息，导致正确率较低，错检率较高。简单地增加网络的层数会导致梯度流消失和信息流弥散等问题，无法有效地训练网络。将密集连接方式引入到反卷积网络中，提出了一种新型的深层密集反卷积神经网络。该网络共有51层卷积权重层，能够自动学习多层级图像的特征，充分挖掘图像信息，并且该网络是端对端可训练的，避免了深层网络中信息传递消失的问题。同时利用反卷积网络实现了像素级别的建筑物提取，在ISPRS 2D的遥感标注数据集上有良好的表现，具有较强的实际应用价值。

关键词: 建筑物提取, 遥感图像, 密集连接, 卷积神经网络, 反卷积