出处 CVPR2017

Motivation

尝试用条件GAN网络来做image translation,让网络自己学习图片到图片的映射函数,而不需要人工定制特征。

Introduction

作者从不同种类的语言翻译类比,提出了Image translation的概念,并希望在给定足够的训练数据以后,训练后的网络能像完成自动语言翻译的任务一样自动地完成图片的转换任务。

《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP

(这图对俺启发挺大啊,嘻嘻)

GAN损失函数:《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP

L1损失函数:《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP用于约束最终生成图片和原图的偏差不至太大

总损失函数:《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP

生成器G采用了Unet结构来做一些layer的skip,保留必要的信息不被变更。

《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP

UNet是借鉴了[1]提出的架构

《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP

判别器D使用了PatchGAN的判定方式,在图像的每个《Image-to-Image Translation with Conditional Adversarial Networks》论文笔记-LMLPHP块上去判断是否为真,最终平均给出结果

源码 https://github.com/phillipi/pix2pix

[1]O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015.

05-11 22:34