Scholarlink

Converting mathematical formulas (MF) from images into LaTeX sequences poses a substantial difficulty, primarily due to the two-dimensional structure of MF and the limited availability of comprehensive training datasets. In this study, we present Sumen model (Scaling Up Image-to-LaTeX Performance), an innovative end-to-end Transformer model equipped with a Swin Transformer encoder and a decoder-Transformer architecture, designed to address these challenges. Utilizing the largest dataset to date, Sumen achieves notable advancements in accuracy metrics such as BLEU (95.59), Edit Distance (97.3), and Exact Match (69.23) on the img2latex100k benchmark, along with significant improvements across CROHME datasets for handwritten MF. These findings highlight Sumen’s effectiveness as a versatile tool for various formula recognition tasks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sumen: End-To-End Transformer Model for Image-To-LaTeX with Enhanced Performance and Large Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bay Nguyen Van,
  • Trung Nguyen Quoc,
  • Vinh Truong Hoang,
  • Viet-Tuan Le,
  • Longfeng Hou

摘要

Converting mathematical formulas (MF) from images into LaTeX sequences poses a substantial difficulty, primarily due to the two-dimensional structure of MF and the limited availability of comprehensive training datasets. In this study, we present Sumen model (Scaling Up Image-to-LaTeX Performance), an innovative end-to-end Transformer model equipped with a Swin Transformer encoder and a decoder-Transformer architecture, designed to address these challenges. Utilizing the largest dataset to date, Sumen achieves notable advancements in accuracy metrics such as BLEU (95.59), Edit Distance (97.3), and Exact Match (69.23) on the img2latex100k benchmark, along with significant improvements across CROHME datasets for handwritten MF. These findings highlight Sumen’s effectiveness as a versatile tool for various formula recognition tasks.