arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2009.13037eess.IV
收藏

混合光谱生成对抗网络在不平衡高光谱图像分类中的应用

Mixture of Spectral Generative Adversarial Networks for Imbalanced Hyperspectral Image Classification

Tanmoy Dam, Sreenatha G. Anavatti, Hussein A. Abbass

摘要:

我们提出了一种三层谱生成对抗网络(GAN)架构,使GAN能够在不平衡条件下管理少数类。一个类依赖的混合生成器频谱GAN(MGSGAN)已被开发,以迫使生成的样本保持在数据的实际分布的域。MGSGAN即使在多数类与少数类的不平衡比率很高的情况下也能够生成少数类。采用基于低层特征的分类器和序贯遗传算法组成三人GAN博弈。生成器网络执行数据增强以提高分类器的性能。该方法已通过两个高光谱图像数据集进行了验证,并与国家的最先进的方法相比,在两个类的不平衡设置对应于真实的数据分布。

英文摘要:

We propose a three-player spectral generative adversarial network (GAN) architecture to afford GAN with the ability to manage minority classes under imbalance conditions. A class-dependent mixture generator spectral GAN (MGSGAN) has been developed to force generated samples remain within the domain of the actual distribution of the data. MGSGAN is able to generate minority classes even when the imbalance ratio of majority to minority classes is high. A classifier based on lower features is adopted with a sequential discriminator to form a three-player GAN game. The generator networks perform data augmentation to improve the classifier's performance. The proposed method has been validated through two hyperspectral images datasets and compared with state-of-the-art methods under two class-imbalance settings corresponding to real data distributions.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑