人工智能培训

搜索

人工智能论文:SIFTINGGAN:生成和筛选标记样本以改善体外遥感图像场景分类基线(SiftingGAN: Generating and Sifting La

[复制链接]
632956335 发表于 2018-9-14 09:12:02 | 显示全部楼层 |阅读模式
632956335 2018-9-14 09:12:02 67 0 显示全部楼层
人工智能论文:SIFTINGGAN:生成和筛选标记样本以改善体外遥感图像场景分类基线(SiftingGAN: Generating and Sifting Labeled Samples to Improve the Remote  Sensing Image Scene Classification Baseline in vitro)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。许多研究试图借助生成对抗网络(GAN)的无监督学习来解决这个问题。然而,在这些研究中,生成的样本仅用于GAN内部进行训练,这些样本尚未证明使用asaugmentation数据训练其他深层网络的GAN生成样本的有效性。此外,传统的图像转换操作(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,问题是GAN产生的样品是否比转化的样品表现更好,需要进行研究。因此,我们提出了一个SiftingGAN框架来生成更多,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,采用在线输出方法生成样本,采用模型筛选的生成模型筛选方法,以及用于采样的Labeled-Sample-Discriminating方法。我们通过改变原始增强数据比率和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由提出的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更好地执行。
Lack of annotated samples vastly restrains the direct application of deeplearning supervised method in remote sensing scene classification.Manyresearches try to tackle this issue with the aid of unsupervised learningability of generative adversarial networks (GANs).However, in theseresearches, the generated samples are only used inside the GANs for training,which haven't proved the effectiveness of the GAN-generated samples using asaugmentation data for training other deep networks.Moreover, traditional imagetransformation operations such as flip and rotation, are still broadly appliedfor data augmentation but limited in quantity and diversity.Thus the questionwhether the GAN-generated samples perform better than the transformed samplesremains to be research.Therefore, we propose a SiftingGAN framework togenerate more numerous, more diverse, more authentic labeled samples for dataaugmentation.SiftingGAN extends traditional GAN framework with anOnline-Output method for sample generation, a Generative-Model-Sifting methodfor model sifting, and a Labeled-Sample-Discriminating method for samplesifting.We conduct three groups of control experiments by changing theoriginal-augmented data ratio and applying different augmented samples.Theexperimental results on AID dataset verify that the samples generated by theproposed SiftingGAN effectively improve the scene classification baseline andperform better than the samples produced by traditional geometrictransformation operations.人工智能论文:SIFTINGGAN:生成和筛选标记样本以改善体外遥感图像场景分类基线(SiftingGAN: Generating and Sifting Labeled Samples to Improve the Remote  Sensing Image Scene Classification Baseline in vitro) dB39X53xUCiU5fih.jpg
URL地址:https://arxiv.org/abs/1809.04985     ----pdf下载地址:https://arxiv.org/pdf/1809.04985    ----人工智能论文:SIFTINGGAN:生成和筛选标记样本以改善体外遥感图像场景分类基线(SiftingGAN: Generating and Sifting Labeled Samples to Improve the Remote  Sensing Image Scene Classification Baseline in vitro)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

632956335当前离线
新手上路

查看:67 | 回复:0

快速回复 返回顶部 返回列表