Few-Shot Writer Identification with Vision Transformers

Prototype-based meta-learning for identifying writers from limited handwriting samples.

Approach

Writer identification usually assumes abundant handwriting data per person. This personal side project targeted the few-shot setting instead: identifying a writer from only a small number of handwriting samples.

  • Extracted global handwriting features with a Vision Transformer backbone.
  • Built per-writer style prototypes with a Prototypical Network so that new writers can be recognized from limited samples.
  • Updated prototypes from user feedback on wrong predictions, incrementally improving accuracy with little data.
  • Built the data and training pipeline in PyTorch.

Vision Transformer · meta-learning · PyTorch · Python

Presentation: 김인곤, 신수용, “소량의 필적 샘플을 활용한 Vision Transformer 기반 메타러닝 작성자 식별 기법,” 한국통신학회 학술대회논문집, pp. 582–583, 2025.

Patent: A Korean patent application covering this method is listed on the Patents page.