# Facial Expression Recognition by De-expression Residue Learning阅读笔记

## Inspriation

1. people are capable of recognizing facial expressions by comparing a subject’s expression with a reference expression (i.e., neutral expression) of the same subject[1].
2. a facial expression can be decomposed to an expressive component and neutral component[2]
人们可以通过一个参考表情来识别其它表情（这里参考表情用的是中性表情）；一个人脸表情可以分为中性部分和表情部分。如图所示。

## Network architecture

### Generator

cGAN[3]被用来从一个给定的图片生成一个中性人脸表情。
cGAN训练的输入是一个图像对，而生成器的输出为I_{output}。其中I_{target}是图片的中性表情的ground truth。I_{output}输出的是GAN生成的中性表情
The discriminator tries to distinguish the from the

the generator tries to not only maximally confuse the discriminator but also generate an image as close to the target image as possible.

Generator的目标函数

Discriminator的目标函数

cGAN的目标函数

CK+与各个方法准确率比较

CK+混淆矩阵