目的:
探查该方法用不同内容的图片来实现SR重建的方法;
摘要:
- 为GANs提出了一种新的判别器(基于Hypersphere embedding space),它能够通过学习超球面空间特征?来提升生成样本的真实性和多样性;
Our work also adopts a hypersphere embedding space, but proposes a different strategy in
structuring and learning the hypersphere, which will be discussed in details.
- Extend the proposed method for conditional settings with class labels by creating a hypersphere for each category and performing class-wise discrimination and update.
关于:spherical circles
Given a unit pivotal vector 𝑝, which is learnable, we can define the set of spherical circles Ωp that is perpendicular to 𝑝.
1、介绍:
- Gradient is difficult to stablize
- Lipschitz continuous space
- Hypersphere embedding space 文章采用该方法
- Prototype(鉴别器基于它来判断真实性的程度)真实性判别指标:仅仅一个是不够的,所以引入了多鉴别器;
- Conditional GANs:使用“Class-Label”,但是不能保证 intra-class diversity,所以仍然存在mode collapse。
- 文章解决的问题:
- 问题:mode collapse(即 lack of diversity)
保证了学习后生成samples的多样性
2、相关工作:
2.1 Generative Adversarial Networks:
多种针对mode collapse的不同解决办法;
2.2 Conditional GANs:
提升GANs效果的直接方法是“利用sample 的 side information”(特别是class labels),根据使用方法不同,可以分为两类:projection-based models、classifier-based models;(作者用的不是这两类方法)
- projection-based models:缺点是do not learn class embeddings through an explicit classifier, it may not be easy for the methods to learn the class-specific features
- classifier-based models:它使用了auxiliary classifier with a discriminator to explicitly learn class-specific features for generation,然而带来了新问题——生成样本的多样性会下降严重;为了解决这个问题引入了adversarial loss on the classifier
作者(为了提升GANs效果)address the mode-collapse issue based on the hypersphere-based discriminators by integrating only a single auxiliary classifier without any adversarial loss.
3、提出方法:
CircleGANs学习了一个鉴别器(它将samples 投射到超球面上,然后用它们对应的超球面spherical circles来打分);(如上图所示)
关键思想是 to leverage the geometric characteristics of the hypersphere in scoring the quality of samples.
First introduce CircleGAN for unconditional settings and then extend it to conditional settings.
3.1 CircleGAN
3.2 Extension to Conditional GANs
The key idea is to create multiple hyperspheres for target categories and perform adversarial learning in a class-wise manner.
(还需要再看)
3.3 Comparison to SphereGAN
对于 Hypersphere projection ,
CircleGAN 采用了 translation 和 l2-Normalization
SphereGAN 使用 inverse stereographic projection(ISP).
最终的效果是circleGAN的projection method要比ISP好。
采用circles被证明对于提升sample的多样性和质量有着很大的帮助;
4、实验:
在unconditional和conditional settings的GANs上都进行试验,来证明该方法的高效性;
框架结构是在 Gans trained by a two time-scale update rule converge to a local nash equilibrium 这篇文章的基础上进行改进的(详见 附件A);
Discriminator 和 Generator 采用的是 ResNet-based 结构;