Scholarlink

<p>The classification of human physical actions is a growing area of study for robotics and human-computer interaction, with the goal of assisting individuals with disabilities in interacting with the real world. The movement of muscle systems involved in human motion is measured by surface electromyography (sEMG) signals, which also provide more information about physical activity. This study proposes a deep learning-based method that utilizes sEMG signals to classify human hand gestures. Time-frequency images (TFIs) of the sEMG signals are created using the smoothed pseudo-wigner-ville distribution (SPWVD), Margenau-hill-spectrogram time-frequency distribution, and Gabor spectrum time-frequency distribution. A ResNet-50 convolutional neural network (CNN) model is then used to extract deep features from the TFI. These deep features are rich in information and essential for accurate classification. But, due to their high dimensionality it can lead to increase computational complexity. To enhance the model’s robustness and computational efficiency, the deep features are reduced to the optimal quantity using the optimal deep feature selection approach (ODFSA). Recursive feature elimination (RFE), principal component analysis (PCA), and mutual information are combined by the ODFSA to enhance model performance and interpretability. The comparison results shows that, the optimal deep feature analysis with the suitable TF method and classifier provides higher accuracy with less computing cost. The best accuracy of 95.4% and 91.3% is achieved for 7 and 10 hand gestures, respectively.</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

Classification of hand gesture based on optimal deep feature selection using sEMG signal

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Akanksha Dixit,
  • Varun Bajaj,
  • Prabin Kumar Padhy

摘要

The classification of human physical actions is a growing area of study for robotics and human-computer interaction, with the goal of assisting individuals with disabilities in interacting with the real world. The movement of muscle systems involved in human motion is measured by surface electromyography (sEMG) signals, which also provide more information about physical activity. This study proposes a deep learning-based method that utilizes sEMG signals to classify human hand gestures. Time-frequency images (TFIs) of the sEMG signals are created using the smoothed pseudo-wigner-ville distribution (SPWVD), Margenau-hill-spectrogram time-frequency distribution, and Gabor spectrum time-frequency distribution. A ResNet-50 convolutional neural network (CNN) model is then used to extract deep features from the TFI. These deep features are rich in information and essential for accurate classification. But, due to their high dimensionality it can lead to increase computational complexity. To enhance the model’s robustness and computational efficiency, the deep features are reduced to the optimal quantity using the optimal deep feature selection approach (ODFSA). Recursive feature elimination (RFE), principal component analysis (PCA), and mutual information are combined by the ODFSA to enhance model performance and interpretability. The comparison results shows that, the optimal deep feature analysis with the suitable TF method and classifier provides higher accuracy with less computing cost. The best accuracy of 95.4% and 91.3% is achieved for 7 and 10 hand gestures, respectively.