深度残差神经网络高分辨率遥感图像建筑物分割

Building Segmentation in High Resolution Remote Sensing Image by Deep ResNet

针对高分辨率遥感图像建筑物分割问题,提出一种Encoder-Decoder的深度学习框架,建立输入图像到分割结果之间的端对端的分割模型。其中Encoder以残差网络为基础,自动提取建筑物的特征;Decoder采用反卷积实现对特征图的上采样,从而完成对建筑物的分割;同时引入批量规范化处理,降低了神经网络权重训练过程中的梯度竞争,从而减小了神经网络的训练难度。实验表明:提出的建筑物分割算法能有效提取建筑物的块状特征和边缘信息,降低复杂道路等干扰的影响,提升建筑物的分割精准度,算法对邻近复杂道路的建筑物、规律性建筑物、单体复杂建筑物等3种典型建筑物的分割精度分别为:0.837、0.892和0.630;F值分别为:0.851、0.879和0.730。同时,多分辨率条件下的分割实验结果表明,该算法对于一定范围内的多分辨率遥感图像具有较好的泛化能力。

This paper addresses the buildings segmentation in high resolution remote sensing image and proposes an Encoder-Decoder architecture of deep learning with End-to-End model,in which Encoder is based on ResNet,and the features needed by segmentation are exacted automatically,and the Decoder produces the segmentation result by deconvolution. Furthermore,in the training process,batch normalization is employed to decrease the gradient competition,so as to reduce the difficulty of training the deep neural network.The experiment results show that the algorithm effectively exacts the bulk feature and edge information of building in the high resolution remote sensing image,therefore the complex road disturbance is suppressed convincingly,and the building segmentation precision is improved effectively,the segmentation precision for three typical buildings,the building besides complex road,the ordered buildings and the complex single building,are 0.836 5,0.892 4,and 0.629 7 respectively;and the F-measure are 0.851 4,0.878 6 and 0.729 8,respectively. Meanwhile,the experiment results for multi-resolution remote sensing images show that the method can be generalized to the multi-resolution image within limits.