Figure 8.
StyleGAN Generator
THE 3
rd
INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS
dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
160
In the recent years, many papers have been published indicating different
methods to manipulate the latent space representations of images used in
StyleGAN architecture by applying certain meaningful transformations.
Primarily there are two kinds of approaches utilized to reconstruct the
decoupled features of images, iterative gradient correction and encoder
architecture based methods. The first is based on repetitive correction of
certain latent features by using gradient backpropagation, like “e2e” encoder
architecture [2]. The latter type of methods are directly applying encoder
network on input image to predict the most similar latent representations, like
Pixel to Style to Pixel architecture which operate globally, allows handling of
non-linear transformations and multi-modal synthesis [3].
My main contributions can be summarized as follows:
I have trained pSp architecture end-to-end on more relevant and angled
face images included in Celeb Faces Attributes dataset and Families In
the Wild (Kinship Recognition) datasets to produce efficient face
encoding & frontalization results.
I have designed new feature mapper Convolutional Neural Network that
will take father and mother images, map them into abstract convolutional
features and apply non-linear transformations to produce child image
features in Ⱳ+ space.
I have performed extensive quantitative and qualitative tests in order to
prove the efficiency of the proposed method and show its power in
kinship image generation.
The method is primarily motivated by ChildGAN architecture which
leverages genetic laws and achieves image generation in two fusion steps
conducted in macro and micro levels [4].
Figure 9.
Pixel to Style to Pixel Network Architecture for Image to Image Translation
|