Scholarlink

<p>Accurate hand gesture recognition plays a vital role in advancing human–computer interaction, particularly in domains such as sign language translation and gesture-based control systems. Despite notable progress, existing methods often struggle in uncontrolled environments, where variations in lighting, background clutter, occlusion, and hand orientation significantly hinder performance. Moreover, conventional approaches typically treat segmentation and classification as independent stages, while isolated models and traditional ensemble schemes often underperform in case of visually similar gestures and the variability of unconstrained environments. This paper introduces DCapNet, an integrated deep learning framework that employs U-Net–based hand segmentation and multi-model classification with a novel confidence-based ensemble mechanism. The segmentation module isolates hand regions from RGB images, and the extracted regions are subsequently fed to four state-of-the-art convolutional neural networks (ResNet50, EfficientNetB0, InceptionV3, and XceptionNet). The ensemble strategy then aggregates the confidence scores from all models to determine the final class corresponding to the highest cumulative score, effectively utilizing the complementary strengths of each architecture while minimizing individual model biases. Performance across several experiments and related ablation studies underscore the superiority of the developed framework, achieving 96.32% accuracy with precision of 0.9640, recall of 0.9632, F-measure of 0.9629, and RMSE of 0.7518 on the OUHANDS dataset. Compared to standalone models and existing relevant state-of-the-art approaches, the framework is found to outperform consistently, and thereby providing a robust and scalable framework for vision-based hand gesture recognition. The source codes and usage guidelines have been made available at <a href="https://github.com/taniyasahana-19/DCapNet">https://github.com/taniyasahana-19/DCapNet</a>.</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 integrated segmentation-classification framework with deep confidence ensemble for hand gesture recognition

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Taniya Sahana,
  • Ayatullah Faruk Mollah

摘要

Accurate hand gesture recognition plays a vital role in advancing human–computer interaction, particularly in domains such as sign language translation and gesture-based control systems. Despite notable progress, existing methods often struggle in uncontrolled environments, where variations in lighting, background clutter, occlusion, and hand orientation significantly hinder performance. Moreover, conventional approaches typically treat segmentation and classification as independent stages, while isolated models and traditional ensemble schemes often underperform in case of visually similar gestures and the variability of unconstrained environments. This paper introduces DCapNet, an integrated deep learning framework that employs U-Net–based hand segmentation and multi-model classification with a novel confidence-based ensemble mechanism. The segmentation module isolates hand regions from RGB images, and the extracted regions are subsequently fed to four state-of-the-art convolutional neural networks (ResNet50, EfficientNetB0, InceptionV3, and XceptionNet). The ensemble strategy then aggregates the confidence scores from all models to determine the final class corresponding to the highest cumulative score, effectively utilizing the complementary strengths of each architecture while minimizing individual model biases. Performance across several experiments and related ablation studies underscore the superiority of the developed framework, achieving 96.32% accuracy with precision of 0.9640, recall of 0.9632, F-measure of 0.9629, and RMSE of 0.7518 on the OUHANDS dataset. Compared to standalone models and existing relevant state-of-the-art approaches, the framework is found to outperform consistently, and thereby providing a robust and scalable framework for vision-based hand gesture recognition. The source codes and usage guidelines have been made available at https://github.com/taniyasahana-19/DCapNet.