Scholarlink

To address key challenges in Tibetan ancient text recognition—significant disparities across types, scarcity of character-level annotations, and insufficient model transfer capability—this study proposes Tib-DTLR, an adapter-based transfer learning framework. We design a Wavelet Transform Pooling (WTP) module that effectively fuses global morphological and local detail features of characters through multi-scale decomposition, and propose a new Wavelet Transform Visual Adapter (WTVA) that implements dynamic cross-domain feature adaptation via a Channel-Spatial Adaptive Feature Fusion strategy (CSAFF). WTVA adapter achieves SOTA with only 2.60% parameter updates, outperforming full fine-tuning by 1.94% on Norbuketaka.

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

Multi-type Tibetan Ancient Book Text Line Recognition Based on Adapter Fine-Tuning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Feixiang Cui,
  • Dorje Tashi,
  • Yong Tso,
  • Nyima Tashi

摘要

To address key challenges in Tibetan ancient text recognition—significant disparities across types, scarcity of character-level annotations, and insufficient model transfer capability—this study proposes Tib-DTLR, an adapter-based transfer learning framework. We design a Wavelet Transform Pooling (WTP) module that effectively fuses global morphological and local detail features of characters through multi-scale decomposition, and propose a new Wavelet Transform Visual Adapter (WTVA) that implements dynamic cross-domain feature adaptation via a Channel-Spatial Adaptive Feature Fusion strategy (CSAFF). WTVA adapter achieves SOTA with only 2.60% parameter updates, outperforming full fine-tuning by 1.94% on Norbuketaka.