对抗性自动编码网络在高光谱异常检测中的应用  

Adversarial autoencoder for hyperspectral anomaly detection

作　　者：杜谦 谢卫莹 DU Qian;XIE Weiying(Electrical and Computer Engineering,Mississippi State University,MS 39762,USA;State Key Laboratory of Integrated Services Networks,Xidian University,Xi'an 710071,China)

机构地区：[1]美国密西西比州立大学电子与计算机工程系,密西西比39762 [2]西安电子科技大学空天地一体化综合业务网全国重点实验室,陕西西安710071

出　　处：《测绘学报》2023年第7期1105-1114,共10页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(62121001,U22B2014);中国科协青年人才托举工程(2020QNRC001)。

摘　　要：自动编码器(autoencoder,AE)是一种典型的生成模型。由于它具有简单的学习过程、良好的收敛能力和无监督的特性而得到了广泛的应用。AE的目标函数仅是输入输出之间的重构误差。为了提高其性能,提出了对抗性自动编码器(adversarial autoencoder,AAE),可以为原始的AE网络提供变分推理输出。本文回顾有关无监督和半监督的AAE模型在高光谱异常检测(hyperspectral anomaly detection,HAD)中的应用。除了在隐层空间中使用对抗性学习外,还可以通过在编码器的输入和解码器的输出之间添加对抗性学习来提高AAE的性能;通过这种方式,改进后的AAE网络可以更专注于学习数据分布而不仅是点对点的数值重建。试验结果表明,利用这些深度学习模型完成HAD任务的想法超越了传统HAD方法的概念,显著提高了检测性能。Autoencoder(AE)is a typical generative model.It has been widely used due to its simple learning process,good ability for convergence,and unsupervised nature.To improve the performance of AE whose objective function is merely input-output reconstruction error,adversarial autoencoder(AAE)has been proposed,which can provide variational inference to the network output.This paper reviews the use of unsupervised and semisupervised AAE in hyperspectral anomaly detection(HAD).The performance of AAE can be improved by adding adversarial learning between the input of the encoder and the output of the decoder,in addition to the adversarial learning in the latent space in the original AAE.In this way,the network can focus more on learning data distribution rather than point-to-point data reconstruction.The idea of using these deep learning models is beyond the concept of traditional HAD methods,and can significantly improve the detection performance,as demonstrated by real data experiments.

关 键 词：高光谱影像 异常检测 自动编码器 对抗性自动编码器 对抗学习 

分 类 号：P237[天文地球—摄影测量与遥感]