Artistic Glyph Image Synthesis via One-Stage Few-Shot Learning

摘要

Automatic generation of artistic glyph images is a challenging task thatattracts many research interests. Previous methods either are specicallydesigned for shape synthesis or focus on texture transfer. In this paper, wepropose a novel model, AGIS-Net, to transfer both shape and texture stylesin one-stage with only a few stylized samples. To achieve this goal, we rstdisentangle the representations for content and style by using two encoders,ensuring the multi-content and multi-style generation. Then we utilize twocollaboratively working decoders to generate the glyph shape image andits texture image simultaneously. In addition, we introduce a local texturerenement loss to further improve the quality of the synthesized textures. Inthis manner, our one-stage model is much more ecient and eective thanother multi-stage stacked methods. We also propose a large-scale datasetwith Chinese glyph images in various shape and texture styles, renderedfrom 35 professional-designed artistic fonts with 7,326 characters and 2,460synthetic artistic fonts with 639 characters, to validate the eectivenessand extendability of our method. Extensive experiments on both English and Chinese artistic glyph image datasets demonstrate the superiority ofour model in generating high-quality stylized glyph images against otherstate-of-the-art methods.