Cycle generative adversarial network for unpaired sketch-to-character translation

Cartoon characters are currently being used in various applications such as comic and cartoon production. The ability to generate a variety of poses and facial expression of cartoon characters from simple sketches of stickfigures can ease the drawing process in production. Previous studies show only...

Full description

Saved in:
Bibliographic Details
Main Author: Alsaati, Leena Zeini J.
Format: Thesis
Language:English
Published: 2019
Subjects:
Online Access:http://eprints.utm.my/id/eprint/96646/1/LeenaZeiniJAlsaatiMSC2019.pdf.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Cartoon characters are currently being used in various applications such as comic and cartoon production. The ability to generate a variety of poses and facial expression of cartoon characters from simple sketches of stickfigures can ease the drawing process in production. Previous studies show only few research focused on the task of sketch to character translation. With low performance of detecting rare pose features and improving rare feature detection has not been significantly studied. The aim of our research is to investigate the capabilities of generative adversarial networks (GANs) in the application of Sketch to Character translation. A wide range of extended GAN versions has been reviewed and in this research, a new dataset collection has been proposed which consists of images of sketches and cartoon characters that are manually drawn. A Cycle GAN has been implemented and its performance against Conditional GAN is compared. Cycle GAN’s cycle consistent loss is the main reason for learning a mapping between the domain of source images and the domain of target images without the need of paired training samples. Cycle GAN has been proven successful in handling a verity of applications in unpaired translation setting. The Conditional GAN has been also proven successful in a wide range of applications, however, it requires paired training samples. Results show that Conditional outperforms the Cycle GAN in accurately mapping the cartoon characters to the stickfigure, which is due to the nature of the paired training sample. However, the Cycle GAN still managed to produce sharper images that compete with the results of a Conditional GAN.