基于半监督生成对抗网络的遥感影像地物语义分割

现有的遥感影像地物提取方法大多是利用人工设定特征或神经网络全监督学习特征检测,前者适用范围较小,后者适用范围较大但需要大量标签。为减少影像标签绘制成本和提高在少量有标签数据下网络的检测精度,本文提出一种新的网络组合构建生成对抗网络,并将其结合半监督学习首次应用到遥感领域进行影像地物检测。文中首先采用选择合适的生成网络和鉴别网络构建生成对抗网络;然后采用有标签数据和无标签数据交替训练网络,根据网络性能选择设置最优参数。本文采用ISPRS提供的vaihingen地区高分辨率航空影像进行实验,结果表明,本文提出的网络组合结合半监督学习可以有效提高检测精度。

Semantic Segmentation of Objects in Remote Sensing Images Based on Semi-supervised Generative Adversarial Nets

Most of the existing remote sensing image feature extraction methods use manual setting features or neural network full-supervised learning feature detection.The former has a small application range,while the latter has a large application range but requires a large number of tags.In order to reduce the cost of image labeling and improve the detection accuracy of the network under a small amount of tagged data,this paper proposes a new network combination to build a confrontation network,and combines it with semi-supervised learning for the first time in the field of remote sensing for image detection.Firstly,the appropriate generation network and the authentication network are used to construct the confrontation network,and then the network is alternately trained with the tagged data and the unlabeled data,and the optimal parameters are set according to the network performance.In this paper,the high-resolution aerial imagery of the vaihingen area provided by ISPRS is used for experiments.The results show that the proposed network combination combined with semi-supervised learning can effectively improve the detection accuracy.