fnctId=thesis,fnctNo=358
[심성현] Legacy Learning Strategy Based on Few-Shot Font Generation Models for Automatic Text Design
- 작성자
- scsc연구센터
- 저자
- Younghiw Kim, Sunghyun Sim, Dohee Kim, seok chan Jeong
- 발행사항
- 발행일
- 20260310
- 저널명
- Information Processing & Management
- 국문초록
- 영문초록
- Text design plays a vital role in enhancing user experience and maintaining visual coherence
across digital platforms. However, for languages containing thousands of glyphs, creating new
fonts remains costly and labor-intensive. Few-shot font generation (FFG) has emerged as a
promising approach that synthesizes unseen glyphs from a small number of reference samples.
Nevertheless, over 5?10% of glyphs must still be manually produced, limiting scalability. To
overcome this limitation, this study proposes a plug-and-play legacy learning strategy that enhances both stylistic diversity and structural stability. The model-agnostic method performs
iterative cross-font fine-tuning, progressively aligning the pretrained models' style representation
toward a legacy-centric embedding space while introducing a content preservation loss to
constrain geometric deformation and maintain glyph integrity. The proposed strategy was applied
to five state-of-the-art FFG models and evaluated on large-scale Korean, Chinese and Thai font
sets. Experimental results based on Fr´echet Inception Distance (FID), Learned Perceptual Image
Patch Similarity (LPIPS), and Centered Kernel Alignment (CKA) similarity show that the generated glyphs gradually diverged from the original font distribution while converging toward the
legacy font distribution, achieving stylistic variety with structural coherence. In addition, quantitative assessments of structural preservation and readability (Dice similarity, Chamfer distance,
and OCR) confirm geometric stability without structural collapse. Cross-language and multilingual generalization was further verified, and a user study with 70 participants (experts and
nonexperts) achieved average usability scores above 4.5/5. These results highlight legacy
learning as an efficient and scalable strategy for automated, high-quality font generation in
multilingual digital content environments.
- 일반텍스트
- 첨부파일