A CycleGAN Framework for Anime Style Image Synthesis based on U-Net and Self-Attention

Abstract

In the contemporary realm of digital imagery, Image Synthesis (IS) is a technique employed to craft artificial images that encapsulate specific content tailored to user preferences. Owing to the intricate and time-intensive nature of this process, researchers have turned to Generative Adversarial Networks (GANs). These networks based on back-propagation signals, alleviate the need for extensive annotated training data. This paper introduces an Animation Style Image Synthesis from natural images by utilizing the CycleGAN framework. Initially, the exploration of CycleGAN’s capabilities focused on style transfer to integrate it with U-Net architecture for the creation of anime-style images. Gradually, enhancements in feature extraction and the improvement in overall image quality has been carried out by incorporating self-attention layers and ResNets. The experimental outcomes for proposed architecture have been verified against established evaluation criteria that indicates a promising direction for research in the field of IS.