A Study on CartoonGAN Using High-Resolution Generative Networks Through Feature Emphasis

GAN is a model that generates completely new data and can create various media contents. CartoonGAN is a GAN model that converts real images into cartoon style, and can apply various cartoon styles. In this study, we propose ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks) technology as a preprocessing method to improve the performance of CartoonGAN. ESRGAN is a model that improves resolution using GAN. This method showed better performance than other SR techniques in the evaluation indices MSE, PSNR, and SSIM results. This suggests that ESRGAN is effective in high-resolution restoration ability and visual quality improvement in cartoon style image conversion. The method proposed in this study can be used as a useful tool for high-resolution cartoon image generation. Since the learning data of the model is limited to a specific domain, there may be limitations in generalizability. Due to the characteristics of ESRGAN, which is computationally complex and consumes a lot of resources, there may be limitations in applying it to real-time processing or large datasets. In future studies, it is necessary to build an extended dataset for various images to overcome these limitations. It is important to reduce computational costs by making models lightweight and optimizing them, and to develop algorithms suitable for real-time processing. It is expected that these additional studies and improvements will be utilized to produce high-quality images.