高分辨率遥感影像语义分割的半监督全卷积网络法

High-resolution remote sensing image semantic segmentation based on semi-supervised full convolution network method

在遥感领域,利用大量的标签影像数据来监督训练全卷积网络,实现影像语义分割的方法会导致标签绘制成本昂贵,而少量标签数据的使用会导致网络性能下降。针对这一问题,本文提出了一种基于半监督全卷积网络的高分辨率遥感影像语义分割方法。通过采用一种集成预测技术,同时优化有标签样本上的标准监督分类损失及无标签数据上的非监督一致性损失,来训练端到端的语义分割网络。为验证方法的有效性,分别使用ISPRS提供的德国Vaihingen地区无人机影像数据集及国产高分一号卫星影像数据进行试验。试验结果表明,与传统方法相比,无标签数据的引入可有效提升语义分割网络的分类精度并可有效降低有标签数据过少对网络学习性能的影响。

In the field of remote sensing,the method of realizing image semantic segmentation by using a large amount of label image data to supervise training full convolution network will result in expensive label drawing cost,while the use of a small amount of label data would lead to network performance degradation.To solve this problem,this paper proposes a semi-supervised full convolution network based semantic segmentation method for high resolution remote sensing images.Specifically,we explore an ensemble prediction technique to train the end-to-end semantic segmentation network by simultaneously optimizing a standard supervised classification loss on labeled samples along with an additional unsupervised consistence loss term imposed on labeled and unlabeled data.In the experiments,the image data set of Vaihingen in Germany provided by ISPRS and satellite GF-1 data were used,and the experimental results show that the proposed method can effectively improve the network performance degradation caused by using only a small amount of label data.