Korean Font Synthesis with GANs

Abstract

Font synthesis for CJK based languages that consists of large number of characters and complex structures is still a major challenge and ongoing research problem for computer vision and AI. In this paper, we propose a generative model based on GANs as a solution for Korean font synthesis problem with a small set of characters. Korean Hangul includes 11,172 characters and composes of writing in multiple patterns. Normally font design involves heavy loaded human labor that can easily hit to one year to finish for one style set. Various methods have been proposed to solve this character generation problem using generative models such as GANs, but the results are often blurry or broken and are far from realistic. We generate visually appealing Korean Hangul characters with a skeleton-driven approach. We demonstrate that this approach is effective at synthesizing characters from their corresponding skeletons. With 114 samples, the proposed method automatically generates the rest of the characters in the same given font style. Our approach resolves long overdue shortfalls such as blurriness, breaking, and unrealistic shapes and styles of characters using GANs. We demonstrate via our experiments that our approach has better quality then other methods.