AnoGAN异常检测
f-AnoGAN异常检测

AnoGAN

Schlegl, Thomas, et al. “Unsupervised anomaly detection with generative adversarial networks to guide marker discovery.International conference on information processing in medical imaging. Springer, Cham, 2017.

  1. DCGAN训练完毕后,冻结DCGAN的参数,隐空间为待训练参数。随机在**隐空间中选择一个点**![](https://cdn.nlark.com/yuque/__latex/33db775ca761534d212472d47c4bd7e4.svg#card=math&code=z_y&height=16&width=15),生成器生成的图像记为![](https://cdn.nlark.com/yuque/__latex/cbdf9caba63e0bbe165c55b8a4cf6cbe.svg#card=math&code=G%28z_y%29&height=21&width=41),测试图像记为x,通过使用反向传播,利用某些优化算法,例如SGD,经过若干轮优化,可以在隐空间中找到一个点,记为![](https://cdn.nlark.com/yuque/__latex/69c9b595a006ad629d9ae24ffad2d385.svg#card=math&code=z%5E%7B%27%7D_y&height=24&width=15)使得loss最小,此时生成器生成的图像![](https://cdn.nlark.com/yuque/__latex/f40c9c1e4adde5ddcd37cf6d58d8f3c1.svg#card=math&code=G%28z%5E%7B%27%7D_y%29&height=24&width=41),与图像A最为相近。由于生成器只能生成正常图片,异常图片与正常图片的差距本来就比较大,比较图像![](https://cdn.nlark.com/yuque/__latex/f40c9c1e4adde5ddcd37cf6d58d8f3c1.svg#card=math&code=G%28z%5E%7B%27%7D_y%29&height=24&width=41)与图像x的差距,**如果差距大于某个数值,则可认为图像A出现异常。整个网络框架如下所示:**<br />![image.png](https://cdn.nlark.com/yuque/0/2020/png/353459/1593489475639-93168863-b737-4a3c-84bd-0596769b37f3.png#align=left&display=inline&height=436&margin=%5Bobject%20Object%5D&name=image.png&originHeight=436&originWidth=901&size=188615&status=done&style=none&width=901)<br />由此可以抛出两个问题
  • 如何设计loss
  • 如何比较比较图像Journey_Competition - 图1与图像x的差异。

Generative Adversarial Representation Learning

Unsupervised Manifold Learning