Dual-Encoding Style Transfer for Korean Font Generation

Abstract

Font generation is time-consuming and labor-intensive, in particular for Korean fonts with 11792 letters in total. Herein, we propose a dual-encoding style transfer network that is specifically designed for Korean font generation. Provided with the letters in a source target, the objective of the network is to learn the style of a target font and to generate letters in the target font. The network consists of a dual-encoder and a decoder. The dual encoder takes a source letter and a combination of one or two consonants and a vowel of a target letter, extracting letter embeddings. The decoder receives the embeddings and generates a letter in the target font. In the learning phase, the network receives 2350 letters and learns the style of the font of interest. Then, the network generates the rest of 8822 letters that the network has not seen during the learning phase. We employ 27 Korean fonts to evaluate the proposed network. The experimental results demonstrate that the proposed network is able to learn the style of various fonts and to generate the full letter set, leading to an efficient and effective design and creation of Korean fonts.