Scholarlink

<p>Keyword spotting (KWS) refers to retrieving all occurrences of a query word from document images without requiring full transcription. While offline handwritten word spotting has achieved notable progress for languages such as English, it remains largely unexplored for South Indian scripts—particularly Malayalam, whose conjunct consonants, ascenders, and variable word boundaries create significant challenges for recognition and retrieval. This work introduces the first end-to-end offline handwritten Malayalam word spotting framework that seamlessly integrates an automatic annotation pipeline with an adaptive triplet-loss network. To the best of our knowledge, this is the first reported system to address Malayalam’s script-specific complexities through a unified annotation and retrieval approach, enabling practical word spotting for a previously unaddressed low-resource script. In the absence of a public benchmark, the system was evaluated on a custom Malayalam handwritten dataset, achieving a mean Average Precision (mAP) of 0.7214 for word spotting, a Word Error Rate (WER) of 0.1084 for recognition, and a manual correction rate below 11% for annotation.</p>

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

An end-to-end malayalam handwritten word spotting framework integrated with automatic annotation

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Anitha Mary M. O. Chacko,
  • D Harikrishnan,
  • V J Manoj

摘要

Keyword spotting (KWS) refers to retrieving all occurrences of a query word from document images without requiring full transcription. While offline handwritten word spotting has achieved notable progress for languages such as English, it remains largely unexplored for South Indian scripts—particularly Malayalam, whose conjunct consonants, ascenders, and variable word boundaries create significant challenges for recognition and retrieval. This work introduces the first end-to-end offline handwritten Malayalam word spotting framework that seamlessly integrates an automatic annotation pipeline with an adaptive triplet-loss network. To the best of our knowledge, this is the first reported system to address Malayalam’s script-specific complexities through a unified annotation and retrieval approach, enabling practical word spotting for a previously unaddressed low-resource script. In the absence of a public benchmark, the system was evaluated on a custom Malayalam handwritten dataset, achieving a mean Average Precision (mAP) of 0.7214 for word spotting, a Word Error Rate (WER) of 0.1084 for recognition, and a manual correction rate below 11% for annotation.