Generate Desired Images from Trained Generative Adversarial Networks

Published in: 2019 International Joint Conference on Neural Networks (IJCNN)

Date of Conference: 14-19 July 2019

Date Added to IEEE Xplore30 September 2019 

IDOI: 10.1109/IJCNN.2019.8851911

Publisher: IEEE

Conference Location: Budapest, Hungary, Hungary

Abstract: The emerging of Generative Adversarial Networks (GANs) gives rise to a significant improvement in image generation. However, a controllable way of synthesizing images with specific characteristics still is a challenging issue. Many existing methods are not efficient enough that require additional information and pre-designed attributes, and are with much more human intervention. In this paper, we propose GAGAN, an extension method to the Generative Adversarial Network, which is the first work to generate specific images from a trained GAN model. To control the characteristics of images, a DNA pool of the trained GAN model is introduced and evolved by a genetic algorithm (GA). Then, with the DNA pool, GAGAN can generate the corresponding latent vector (DNA) of target images. Furthermore, GAGAN can synthesize images containing a single specific characteristic or multiple specific attributes (including AND and OR relation). Moreover, several fitness evaluation strategies are also proposed to make GAGAN flexible to control the target characteristics. Experiments on CelebA and MNIST are conducted, and results show that the proposed method is feasible and effective in specific image generation problem.

Share This Page:

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.