Automated image generation

• Main Author: Goh, Shan Ying
• Other Authors: Lu Shijian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/171948
• Institution: Nanyang Technological University

Image translation techniques have gained significant attention in recent years, particularly CycleGAN. Traditionally, building image-to-image translation models requires the collection of extensive datasets with paired examples, which can be complicated and costly. However, CycleGAN’s automatic training approach eliminates the need for such paired samples, thus simplifying the training process while enhancing the potential of image translation, allowing for imaginative and lifelike adjustments. For instance, CycleGAN can effortlessly transform styles like cats into dogs and vice versa, extending to practical domains like art, fashion, and medical imaging. Nevertheless, CycleGAN's applicability in real-world scenarios is limited by its current constraint to a relatively small set of available styles. This compels us to explore more practical alternatives. This study introduces new styles into the framework, assessing their practical effectiveness and addressing concerns about potential loss in image quality. Results show the promising potential of these improved CycleGAN variants for various domains and applications. Keywords: CycleGAN, Style Transfer, Image Translation, Diverse Aesthetics, Creative Applications, Content Preservation