Scholarlink

<p>The development of Bengali Handwritten Text Recognizer (HTR) remains a domain of active research and development due to its script-specific challenges. Although a substantial amount of work has been done and several traditional and deep learning models have been proposed, most of these methods delivered suboptimal accuracy when evaluated on student-written pages collected from Bengali medium schools. This is probably because the training and the testing datasets are different. Thus, this paper aims to design a Bengali HTR system that provides better performance in a cross-domain testing scenario. For that, a Convolutional Neural Network (CNN) with a Transformer-based model is proposed. The CNN-Transformer model achieved a character recognition rate (CRR) of 95.18% when tested using a left-aside portion from the same dataset and a CRR of 89.48% on cross-dataset evaluation. The proposed model is also superior to the existing models tested using the same cross-dataset setting.</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

A CNN-Transformer Approach for Bengali Handwritten Word Recognition

  • 期刊论文

  • 发表日期：

    2025年6月14日

查看全文
  • Sujan Kumar Saha,
  • Vishal Bania,
  • Moumita Moitra

摘要

The development of Bengali Handwritten Text Recognizer (HTR) remains a domain of active research and development due to its script-specific challenges. Although a substantial amount of work has been done and several traditional and deep learning models have been proposed, most of these methods delivered suboptimal accuracy when evaluated on student-written pages collected from Bengali medium schools. This is probably because the training and the testing datasets are different. Thus, this paper aims to design a Bengali HTR system that provides better performance in a cross-domain testing scenario. For that, a Convolutional Neural Network (CNN) with a Transformer-based model is proposed. The CNN-Transformer model achieved a character recognition rate (CRR) of 95.18% when tested using a left-aside portion from the same dataset and a CRR of 89.48% on cross-dataset evaluation. The proposed model is also superior to the existing models tested using the same cross-dataset setting.