基于改进型U-Net网络的高分辨率遥感影像建筑物提取

针对传统人工提取方法自动化程度低、过分依赖人工设计的特征,以及现有的深度学习方法中存在的提取精度不高等问题,提出了一种基于改进型U-Net网络的高分辨率遥感影像建筑物提取方法。首先将空洞卷积加入到网络中,利用不同尺度的空洞卷积对来自网络编码部分的结果进行多尺度特征提取;再对提取的特征进行特征融合,并输入到网络的下一层;然后将制作的数据集输入到网络中进行训练;最后利用Softmax得到最终分割结果。在建筑物公开的数据集中进行测试,提取结果的像素精度为96.26%;Iou精度为78.59%、Recall为95.65%,表明该方法具有良好的鲁棒性和精度,能从影像中准确地提取建筑物。

Building Extraction Method in High-resolution Remote Sensing Images Based on Improved U-Net Network

The traditional manual extraction method has low automation,overreliance on artificial design,and the extraction accuracy of existing extraction methods of deep learning methods are not high.On the basis of this,we proposed a building extraction method in high-resolution remote sensing images based on improved U-Net network.Firstly,we added the cavity convolution into the network,and used several different scale convolutions to extract multi-scale feature for the results from the network coding part.Then,we fused the extracted features,input to the next layer of the network,and inputted the produced data set into the network for training.Finally,we used Softmax to obtain the final segmentation result,and conducted the test in the data set published by the building.The results show that the pixel accuracy of extraction result is 96.26 %,the accuracy of Iou is 78.59%,and the Recall is 95.65 %,which indicates that the proposed method has good robustness and precision,and can accurately extract buildings from images.