用深度学习做人脸合成,website:https://zo7.github.io/blog/2016/09/25/generating-faces.html
受启发于 Learning to Generate Chairs, Tables, and Cars with Convolutional Networks
模型描述
给定一个数据集
包含:c – the one-hot encoding of the model identity
v – azimuth and elevation of the camera position
Θ – the parameters of additional artificial transformations (增加训练集的数量,减少过拟合)
目标(the RGB output image x, the segmentation mask s)
网络结构
“1s-S-deep” model
生成网络模型由两阶段构成:
1. FC-1 to FC-4 建立一个分享的、高维的隐表达 h(c,v,Θ)
2. FC-5 and uconv-1 to uconv-4 (这部分定义为u)生成outputimage和segmentation mask
这个 deconvolution network类似于 here, here, or here,首先upsample输入,然后convolution。
该模型建立在Keras上。
网络训练
网络参数W
LRGB(squared Euclidean)和Lsegm(squared Euclidean/negative log-likelihood)是损失函数
用更理论的方法生成新模型,训练一个概率生成模型(FC-2)隐状态z:潜在的椅子图像集合
定义 a segmentation mask si under transformation TΘi
定义the pixels in an image xi
log likelihood of an image and its segmentation mask
网络分析
activating neurons of FC-1 and FC-2 feature maps 见下图(最左边是 setting all neurons of the layer
to zero,其余图像是activating one randomly selected neuron) 并没有太大变化
activating neurons of FC-3 and FC-4 feature maps ,出现视角和类的变化
Images generated from single neurons of the convolutional layers (From top to bottom: uconv-2,
uconv-1, FC-5 of the RGB stream)
接下来,将通过程序进一步理解该模型。