1 INTRODUCTION
本文是CVPR18的扩展
3 METHOD
3.1 Overview

loss包括the traditional squared Euclidean loss、the GAN loss、the identity loss
结构上,判别器是pyramid-structured discriminator
3.2 Generator
生成器是Encoder-Decoder结构,接收input young face x和target age label(or age range)c作为输入,首先使用卷积层转换到latent space,然后连接4个residual blocks,最后连接3个反卷积层,生成age progression result y,即G(x,c)→y
卷积层之后是Instance Normalization和ReLU,最后一层是total variation regularization layer
3.3 Adversarial Learning
3.3.1 Aging Modeling
定义young faces的分布为x∼Pyoung,generated faces的分布为G(x,c)∼PG,target age faces的分布为Pold,我们希望PG=Pold
在原版GAN的优化目标中,判别器的损失函数如下
LGAND=−Ex∼Pyoung,clog[1−D(G(x,c))](2)−Ex∼Poldlog[D(x)]
因为JS divergence is locally saturated,所以当D训练得太好时,造成优化G时梯度消失,本文使用Least Squares GAN的版本
对于D,将actual young faces和generated age-progressed faces作为阜样本,将true elderly face of age range c作为正样本,于是G和D的优化目标分别为
LGAN_G=Ex∼Pyoung,c H(1,D(ϕage(G(x,c))))(3)
LGAN_D=Ex∼Pyoung,y∼Pold,c H([1,0,0],[D(ϕage(y)),D(ϕage(G(x,c))),D(ϕage(x))])(4)
其中,H表示least squares distance,ϕage是提取extract age-related features的网络,预先训练一个VGG16进行age classification,然后去掉FC layer
ϕage提取的特征包括第2、4、7、10层的feature map,与之对应,D有4个分支,每一个分支的输出都是3x3,最终拼接成12x3
D的结构是Conv-BN-LeakyReLU
3.3.2 Progressive Aging Modeling

原始的框架如Figure 4(a)所示
比较常见的做法是对D加一个auxiliary classifier,增加the age classification loss Lage,如Figure 4(b)所示
3.4 Identity Preservation
采用the network of deep face descriptor,记为ϕid
the identity loss定义如下
Lidentity=Ex∼Pyoung,c d(ϕid(x),ϕid(G(x,c)))(9)
其中d表示squared Euclidean distance
3.5 Objective
为了缩小生成图像与原图之间的gap(保证颜色相近),添加一项pixel-wise L2 loss
Lpixel=Ex∼Pyoung,c W×H×C1∥G(x,c)−x∥22(10)
借鉴文献[38]中的total variation regularizer loss,增加一项Ltv,保证spatial smoothness
最终整个framework的训练目标如下
LG=λai∑LGAN_Gi+λpLpixel+λiLidentity+λtLtv(11)
LDi=LGAN_Di(12)