Generating Faces with deconvolution Networks

Using deep learning to face synthesis, website:https://zo7.github.io/blog/2016/09/25/generating-faces.html

Inspired by learning to Generate chairs, Tables, and Cars with convolutional Networks

Model Description

Given a data set

Contains:c , Haven One-hot encoding of the model identity

v –azimuth and elevation of the camera position

Θ – The parameters of additional artificial transformations (increase the number of training sets, reduce overfitting)

Target (the RGB output image x, the segmentation mask s)

Network Structure

"1s-s-deep" model

The build network model consists of two phases:

1. FC-1 to FC-4 Create a shared, high-dimensional implicit expression H (c,v,θ)

2. FC-5 and Uconv-1 to uconv-4 (defined as U) generate outputimage and segmentation mask

This deconvolution network is similar to here, here, or here, first upsample input, then convolution.

The model is built on the Keras.

Network Training

Network Parameters W

Lrgb (squared Euclidean) and lsegm (squared euclidean/negative log-likelihood) are loss functions

Generating new models in a more theoretical way, training a probabilistic generation model (FC-2) hidden State Z: Potential Chair Image Collection

Define a segmentation mask si under Transformation tθi

Define the pixels in an image XI

Log likelihood of an image and its segmentation mask

Network Analysis

Activating neurons of FC-1 and FC-2 feature maps see (leftmost is setting all neurons of the layer

To zero, the rest of the images are activating one randomly selected neuron) and not much changed

Activating neurons of FC-3 and FC-4 feature maps, with changes in perspective and class

Images generated from single neurons of the convolutional layers (from top to Bottom:uconv-2,

Uconv-1, FC-5 of the RGB stream)

Next, the model will be further understood through the program.

Generating Faces with deconvolution Networks