Abstract
Lack of annotated samples vastly restrains the direct application of deep learning supervised method in remote sensing scene classification. Many researches try to tackle this issue with the aid of unsupervised learning ability of generative adversarial networks (GANs). However, in these researches, the generated samples are only used inside the GANs for training, which haven't proved the effectiveness of the GAN-generated samples using as augmentation data for training other deep networks. Moreover, traditional image transformation operations such as flip and rotation, are still broadly applied for data augmentation but limited in quantity and diversity. Thus the question whether the GAN-generated samples perform better than the transformed samples remains to be research. Therefore, we propose a SiftingGAN framework to generate more numerous, more diverse, more authentic labeled samples for data augmentation. SiftingGAN extends traditional GAN framework with an Online-Output method for sample generation, a Generative-Model-Sifting method for model sifting, and a Labeled-Sample-Discriminating method for sample sifting. We conduct three groups of control experiments by changing the original-augmented data ratio and applying different augmented samples. The experimental results on AID dataset verify that the samples generated by the proposed SiftingGAN effectively improve the scene classification baseline and perform better than the samples produced by traditional geometric transformation operations.
Abstract (translated)
缺乏注释样本极大地限制了深度学习监督方法在遥感场景分类中的直接应用。许多研究试图利用生成对抗网络（GAN）的无监督学习能力来解决这个问题。然而，在这些研究中，生成的样本仅用于GAN内部进行训练，这些样本未证明GAN生成的样本使用增强数据来训练其他深层网络的有效性。此外，传统的图像变换操作，例如翻转和旋转，仍广泛应用于数据增加，但数量和多样性有限。因此，GAN产生的样品是否比转化的样品表现更好的问题仍有待研究。因此，我们提出了一个SiftingGAN框架，用于生成更多，更多样化，更真实的标记样本以进行数据增强。 SiftingGAN扩展了传统的GAN框架，其中包括用于样本生成的在线输出方法，用于模型筛选的Generative-Model-Sifting方法，以及用于样本筛选的Labeled-Sample-Discriminating方法。我们通过改变原始增强数据比率并应用不同的增强样本来进行三组对照实验。 AID数据集上的实验结果验证了所提出的SiftingGAN生成的样本有效地改善了场景分类基线，并且比传统几何变换操作产生的样本表现更好。
URL
https://arxiv.org/abs/1809.04985
