Scholarlink

<p>Handwritten character recognition for Indian regional languages remains relatively underexplored, especially for complex scripts such as Odia and Bangla. The challenges primarily arise from intricate character structures and the limited availability of large annotated datasets. Conventional methods based on handcrafted features require significant domain expertise and often lack robustness across diverse handwriting styles. To address these limitations, we present a lightweight convolutional neural network optimized using Particle Swarm Optimization. The optimization process fine-tunes key hyperparameters, including convolutional filters, dropout rate, and learning rate, ensuring improved model performance. A series of preprocessing steps are applied to suppress noise and retain essential structural details of handwritten samples. The proposed approach is evaluated on benchmark datasets: ISI Kolkata numerals and NIT Rourkela OHCS1.0 characters for Odia, and cMATERdb 3.1.1 numerals and 3.1.2 characters for Bangla. Results from experiments demonstrate that the model attains competitive accuracy and outperforms several existing techniques, demonstrating its effectiveness in recognizing regional scripts while maintaining computational efficiency.</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

Lightweight CNN with Particle Swarm Optimization for Odia and Bangla Handwritten Script Recognition

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Pragnya Ranjan Dash,
  • Rakesh Chandra Balabantaray

摘要

Handwritten character recognition for Indian regional languages remains relatively underexplored, especially for complex scripts such as Odia and Bangla. The challenges primarily arise from intricate character structures and the limited availability of large annotated datasets. Conventional methods based on handcrafted features require significant domain expertise and often lack robustness across diverse handwriting styles. To address these limitations, we present a lightweight convolutional neural network optimized using Particle Swarm Optimization. The optimization process fine-tunes key hyperparameters, including convolutional filters, dropout rate, and learning rate, ensuring improved model performance. A series of preprocessing steps are applied to suppress noise and retain essential structural details of handwritten samples. The proposed approach is evaluated on benchmark datasets: ISI Kolkata numerals and NIT Rourkela OHCS1.0 characters for Odia, and cMATERdb 3.1.1 numerals and 3.1.2 characters for Bangla. Results from experiments demonstrate that the model attains competitive accuracy and outperforms several existing techniques, demonstrating its effectiveness in recognizing regional scripts while maintaining computational efficiency.