一种改进U-Net的遥感影像建筑物提取方法

An improved U-Net method for extracting buildings from remote sensing images

针对高分辨率遥感影像信息复杂浅层网络难以对其目标物特征信息充分学习,图像因裁剪导致边缘信息损失使得模型对图像边缘预测效果较差的问题,该文将U-Net收缩路径加深以增强网络对特征信息的学习能力,并加入随机失活函数(Dropout)层抑制过拟合现象的发生,扩张路径中加入批量归一化层以提高网络训练速度,并将忽略边缘交叉熵函数与骰子函数结合构建联合损失函数作为本文模型的损失函数以提高模型对图像边缘的预测效果。实验结果表明:该文方法对建筑物边缘能够进行有效预测;对建筑物轮廓以及较小建筑物的提取较之SVM、主干网络为VGG的U-Net提取效果有所提高;并在应用扩展研究数据集中有着较好的表现。

To address the problems that the complex shallow network of high-resolution remote sensing image information is difficult to fully learn the target feature information,and the loss of edge information due to image cropping makes the model less effective in predicting the edge of the images,the U-Net shrinkage path was deepened to enhance the network’s ability to learn feature information in this paper,and added Dropout layer to suppress the occurrence of over-fitting phenomenon,the batch normalization layer was added to the expansion path to improve the network training speed,and the joint loss function was constructed by combining the neglected edge cross-entropy function with the dice function as the loss function of the model in this paper to improve the prediction effect of the model on image edges.The experimental results show that this method can effectively predict building edges;the extraction of building contours and smaller buildings is improved compared with SVM and U-Net extraction with VGG as the backbone network;and extension in the application research data set has a better performance.