基于BAU-Net的高分辨率遥感图像中建筑物的分割提取

BAU-Net for buildings segmentation in high resolution remote sensing images

高分辨率遥感图像中建筑物的分割提取是现代数字化城市建设、经济预测、国土资源勘探以及国防安全等领域中重要的技术手段。本文基于U-Net网络提出了一种双U型Encoder-Decoder架构的边缘意识U型深度神经网络—BAU-Net。该网络模型首先利用粗特征提取网络对建筑物进行初次识别提取,再通过残差结构的细特征提取网络修正边缘轮廓。通过可行性实验以及与经典特征提取网络对比,本文所采用的模型算法可以在一定程度上克服道路、车辆、树木遮挡以及建筑物阴影干扰,准确有效地识别建筑物主体,精确完整地分割建筑物边缘轮廓。在测试集上准确率达到90.48%、召回率达到91.30%、F值达到90.58%、平均绝对误差达到0.0325,均优于经典分割算法。

The segmentation of buildings in high resolution remote sensing images is an important technical issue in the construction of modern digital city,economic prediction,resources exploration and the national defense security.Based on U-Net,a double U-Net encoder-decoder framework of boundary awareness is proposed.In the model,the coarse feature extraction network is used for the first time to recognize and extract the buildings,and then the edge contour is modified by the refined feature extraction network of the residual structure.Through valid experiments and comparison with some classic feature extraction networks,the results show that the algorithm adopted in this paper could overcome the disturbance produced by roads,vehicles,trees and building shadow,effectively recognize the main part of the buildings,precisely and completely segment the edge of the buildings.And the precision,recall,F-measure and MAE of the test set are 90.48%,91.30%,90.58%and 0.0325 respectively,all of which are superior to the classical segmentation algorithms.