【论文研读】GAN 天气迁移 Weather GAN: Multi-Domain Weather Translation Using Generative Adversarial Networks

从工作里爱了生命,就是通彻了生命最深的秘密。


看到分享一篇用GAN来完成天气图像的迁移,作者认为 In this paper, a new task is proposed, namely, weather translation, which refers to transferring weather conditions of the image from one category to another. 但似乎以前看到过类似天气转换的任务,换了电脑论文都找不到了,如果大家也看到过可以踊跃分享一下。
作者认为weather-cues很重要,用一个分割模块来代表weather-cues,结合一个初始化模块和一个注意力模块,最后结论效果超越了SOTA。

Weather GAN: Multi-Domain Weather Translation Using Generative Adversarial Networks
Xuelong Li, Fellow, IEEE, Kai Kou, and Bin Zhao

文章地址


本文希望完成不同天气情况下图像的转换,直观上来讲我觉得雪天<–>晴天的转换会比较难,因为想做出雪天的纹理会很困难。

作者在生成器里安排了三块操作,用初始化模块来绘制整体框架,用一个注意力模块来找兴趣点,文章的狠活是提出的weather-cues,作者认为weather-cues 是决定天气的真正因素。作者似乎没有安排模块关心图像质量,三个模块苦苦支撑,保证了输出结果图的基本可靠,但图像质量比较一般。

按照作者的规划,他用一个分割模块来找weather-cues,如果从逻辑上讲这样很奇怪,这个weather-cues 代表的是一个兴趣区域吗?如果直观感觉,weather-cues 似乎应该是一些比较微小的细节特征,但对于本任务来讲,很难找到pair数据,所以只能无监督学习,用real image 做一个 reference。这样的话模型很难知道自己在学什么,用一个标定的区域可以一定程度帮助到模型。

在这里插入图片描述
文中说weather-cues 的目标是直接人为标定出来的,如下图所示。这样的话weather-cues 似乎变成了一个天空标注,整个模型的逻辑变为找天空区域。

在这里插入图片描述

从结果看效果还是很好的,给出的失败例子似乎也是因为天空区域比较少。
在这里插入图片描述

总结

结果还有一定提升空间

  • 3
    点赞
  • 4
    收藏
    觉得还不错? 一键收藏
  • 1
    评论
Chapter 1, Introduction to Deep Learning, speaks all about refreshing general concepts and terminology associated with deep learning in a simple way without too much math and equations. Also, it will show how deep learning network has evolved throughout the years and how they are making an inroad in the unsupervised domain with the emergence of generative models. Chapter 2, Unsupervised Learning with GAN, shows how Generative Adversarial Networks work and speaks about the building blocks of GANs. It will show how deep learning networks can be used on semi-supervised domains, and how you can apply them to image generation and creativity. GANs are hard to train. This chapter looks at some techniques to improve the training/learning process. Chapter 3, Transfer Image Style Across Various Domains, speaks about being very creative with simple but powerful CGAN and CycleGAN models. It explains the use of Conditional GAN to create images based on certain characteristics or conditions. This chapter also discusses how to overcome model collapse problems by stabilizing your network training using BEGAN. And finally, it covers transferring styles across different domains (apple to orange; horse to zebra) using CycleGAN. Chapter 4, Building Realistic Images from Your Text, presents the latest approach of stacking Generative Adversarial Networks into multiple stages to decompose the problem of text to image synthesis into two more manageable subproblems with StackGAN. The chapter also shows how DiscoGAN successfully transfers styles across multiple domains to generate output images of handbags from the given input of shoe images or to perform gender transformations of celebrity images. Chapter 5, Using Various Generative Models to Generate Images, introduces the concept of a pretrained model and discusses techniques for running deep learning and generative models over large distributed systems using Apache Spark. We will then enhance the resolution of low quality images using pr

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论 1
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值