Scholarlink

<p>Optical Character Recognition (OCR) has evolved considerably, with modern systems achieving high accuracy for printed text. However, in multilingual countries like India, handwritten character recognition remains challenging due to script diversity and writing variability. This study presents a hybrid feature-based framework for offline handwritten character recognition across three scripts—Devanagari, Gurmukhi, and Roman. The core contribution lies in the systematic design and evaluation of three handcrafted feature types—diagonal, directional, and transition—combined with Support Vector Machine (SVM) classifiers using different kernels. A balanced dataset of 4,850 handwritten samples from fifty writers was compiled to reflect realistic low-resource multilingual scenarios. Among all configurations, diagonal features with the polynomial kernel achieved the highest accuracy of 92.16%, with 5-fold cross-validation confirming stability (± 1.12%). Expanded baselines with modern CNN architectures (ResNet-18 and DenseNet-121) achieved slightly higher accuracy (~ 93%), but at significantly higher computational cost, highlighting the competitiveness of handcrafted approaches in resource-constrained settings. To enhance interpretability, visual demonstrations of character samples, feature extractions, confusion matrices, and error analyses are included. These results demonstrate that carefully engineered handcrafted features remain viable for multilingual OCR and provide a strong benchmark for future hybrid and deep learning models.</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

Hybrid Feature-Based SVM Framework for Multilingual Offline Handwritten Character Recognition

  • 期刊论文

  • 发表日期：

    2026年3月20日

查看全文
  • Horiya Aldeeb,
  • Halijah Hassan,
  • K. N. Raja Praveen,
  • Mandeep Kaur Chohan,
  • Shivakrishna Dasi,
  • Ankur Srivastava,
  • Saurabh Namdev,
  • Ahmed Alkhayyat

摘要

Optical Character Recognition (OCR) has evolved considerably, with modern systems achieving high accuracy for printed text. However, in multilingual countries like India, handwritten character recognition remains challenging due to script diversity and writing variability. This study presents a hybrid feature-based framework for offline handwritten character recognition across three scripts—Devanagari, Gurmukhi, and Roman. The core contribution lies in the systematic design and evaluation of three handcrafted feature types—diagonal, directional, and transition—combined with Support Vector Machine (SVM) classifiers using different kernels. A balanced dataset of 4,850 handwritten samples from fifty writers was compiled to reflect realistic low-resource multilingual scenarios. Among all configurations, diagonal features with the polynomial kernel achieved the highest accuracy of 92.16%, with 5-fold cross-validation confirming stability (± 1.12%). Expanded baselines with modern CNN architectures (ResNet-18 and DenseNet-121) achieved slightly higher accuracy (~ 93%), but at significantly higher computational cost, highlighting the competitiveness of handcrafted approaches in resource-constrained settings. To enhance interpretability, visual demonstrations of character samples, feature extractions, confusion matrices, and error analyses are included. These results demonstrate that carefully engineered handcrafted features remain viable for multilingual OCR and provide a strong benchmark for future hybrid and deep learning models.