Automatic Font Generation Algorithm Based on Image Kansei

Abstract

In this paper, we propose an automatic font generation algorithm which can accurately represent Kansei of the input image. Provided an arbitrary image, a corresponding font is generated according to the emotion information included in the image. The proposed method consists of two stages: learning stage and generation stage. In the first stage, the learning unit learns multiple fonts by using the image generation model “zi2zi” (meaning characters to characters), which is based on the GAN principle. In the generation stage, the score of the emotion information is obtained using “Emotion API” for the analysis of the image, then an inference is made by using the weights of the Generator model obtained by the learning, and the font is generated finally. Evaluation experiments via web questionnaires show that the fonts generated by the proposed algorithm correctly reflect emotions, demonstrating the effectiveness of our algorithm.