Arakawa Kaoru
Department Undergraduate School , School of Interdisciplinary Mathematical Sciences Position Professor |
|
Language | English |
Publication Date | 2021/05 |
Type | International conference proceedings |
Peer Review | Peer reviewed |
Title | Object Design System by Interactive Evolutionary Computation Using GAN with Contour Images |
Contribution Type | Co-authored (other than first author) |
Journal | Human Centered Intelligent Systems |
Journal Type | Another Country |
Publisher | Springer Nature |
Volume, Issue, Page | pp.66-75 |
Total page number | 242 |
Authorship | Corresponding author |
Author and coauthor | Chen Xin, Kaoru Arakawa |
Details | This paper proposes a method to design objects by interactive evolutionary computation (IEC) using generative adversarial network (GAN) with the contour image of the object which the user wants to design. Here, a conditional GAN is adopted in order to utilize the contour image. GANs can generate lifelike images of objects, however, it is hard to control the output, since GAN just generates images randomly with the distribution learned from training data. Thus, IEC is introduced here to control the latent vectors in the conditional GAN. Moreover, using the contour image, conditional GAN can efficiently generate images of objects which the user prefers. In the proposed method the latent vectors in the conditional GAN are optimized for each user, so that the design is satisfactory to him/her through the process of IEC. Experiment results show the high performance of this method. |