Authors:
Fumiya Yamashita
;
Ryohei Orihara
;
Yuichi Sei
;
Yasuyuki Tahara
and
Akihiko Ohsuga
Affiliation:
University of Electro-Communications, Japan
Keyword(s):
Deep Learning, Domain Transfer, Generative Adversarial Network, Unsupervised Learning.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Bioinformatics
;
Biomedical Engineering
;
Soft Computing
;
Visualization
Abstract:
With the development of deep learning, image translation has made it possible to output more realistic and
highly accurate images. Especially, with the advent of Generative Adversarial Network (GAN), it became possible
to perform general purpose learning in various image translation tasks such as “drawings to paintings”,
“male to female” and “day to night”. In recent works, several models have been proposed that can do unsupervised
learning which does not require an explicit pair of source domain image and target domain image, which
is conventionally required for image translation. Two models called “CycleGAN” and “DiscoGAN” have appeared
as state-of-the-art models in unsupervised learning-based image translation and succeeded in creating
more realistic and highly accurate images. These models share the same network architecture, although there
are differences in detailed parameter settings and learning algorithms. (in this paper we will collectively refer
to them as “learning techni
ques”) Both models can do similar translation tasks, but it turned out that there is
a large difference in translation accuracy between particular image domains. In this study, we analyzed differences
in learning techniques of these models and investigated which learning techniques affect translation
accuracy. As a result, it was found that the difference in the size of the feature map, which is the input for the
image creation, affects the accuracy.
(More)