Scholarlink

<p>Sign language recognition is essential for enabling effective communication between deaf and hearing individuals. In this paper, we propose ViT-S-BDSL, a novel model for Bangla Sign Language (BDSL) recognition. Previous studies on BDSL recognition have primarily utilized Convolutional Neural Networks (CNNs). While researchers have actively explored this field, the potential of Vision Transformers (ViT) for BDSL recognition remains largely unexamined. Notably, no documented research has specifically investigated the use of Distillation with No Labels version 2 (DINOv2) ViT-based models for recognizing Bangla Sign Language characters and numerals, highlighting a significant gap in the literature. This work focuses on balancing the complexity and efficiency of a self-supervised, DINOv2-pretrained ViT-S backbone. Leveraging transfer learning, the model is fine-tuned on the well-curated and publicly available BDSL49 dataset, which comprises 14,745 images representing 49 distinct characters and numerals. This fine-tuning process enhances the model’s ability to extract domain-specific, rich visual features from hand signs and perform recognition using a simple linear classifier. The proposed model achieves an accuracy of 99.7% and an average F1-score of 99.7%, outperforming several CNN-based models. ViT-S-BDSL strikes a balance between computational efficiency and robust feature extraction, making it well-suited for domain-specific tasks such as Bangla Sign Language recognition.</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 vision transformer-based fine-tuned DINOv2 model for bangla sign language recognition

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Syeda Anika Tasnim,
  • Rifath Mahmud,
  • Tanvir Ahmed,
  • Debajyoti Karmaker

摘要

Sign language recognition is essential for enabling effective communication between deaf and hearing individuals. In this paper, we propose ViT-S-BDSL, a novel model for Bangla Sign Language (BDSL) recognition. Previous studies on BDSL recognition have primarily utilized Convolutional Neural Networks (CNNs). While researchers have actively explored this field, the potential of Vision Transformers (ViT) for BDSL recognition remains largely unexamined. Notably, no documented research has specifically investigated the use of Distillation with No Labels version 2 (DINOv2) ViT-based models for recognizing Bangla Sign Language characters and numerals, highlighting a significant gap in the literature. This work focuses on balancing the complexity and efficiency of a self-supervised, DINOv2-pretrained ViT-S backbone. Leveraging transfer learning, the model is fine-tuned on the well-curated and publicly available BDSL49 dataset, which comprises 14,745 images representing 49 distinct characters and numerals. This fine-tuning process enhances the model’s ability to extract domain-specific, rich visual features from hand signs and perform recognition using a simple linear classifier. The proposed model achieves an accuracy of 99.7% and an average F1-score of 99.7%, outperforming several CNN-based models. ViT-S-BDSL strikes a balance between computational efficiency and robust feature extraction, making it well-suited for domain-specific tasks such as Bangla Sign Language recognition.