Glyph Generation for Japanese Calligraphy based on Encoding both Content and Style

We propose a method for generating Japanese handwritten calligraphy works. The main component in the proposed method is a generator that extracts both content and style features and generates glyphs using the features. The generator consists of three components: a content encoder, a style encoder, and a mixer. We utilize SimSiam and GAN for training the generator. We pre-train the content encoder and the style encoder with SimSiam to obtain a suitable feature representation. We bring these pre-trained encoders to the generator. We use GAN to train the generator to obtain high-quality generated images. We train the generator on datasets consisting of 41 styles, 6,523 content and 147,564 glyphs. These datasets consist of Japanese calligraphy data, kuzushiji data, and Japanese fonts data. We make qualitative and quantitative comparisons between the proposed model and the baseline model. In qualitative comparisons, we generate and compare three styles in our proposed and baseline models. In the quantitative comparison, we compare the proposed model and the baseline model with three evaluation metrics: MAE, SSIM, and LPIPS. MAE and SSIM measure similarity at the pixel level. LPIPS quantifies the perceptual similarity. Our proposed model scores MAE 0.2096, SSIM 0.5870, and LPIPS 0.2214, which are 0.0212, 0.0390, and 0.0107 points superior to the baseline, respectively.