Feature recognition in multiple CNNs using sEMG images from a prototype comfort test.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Surface electromyographic (sEMG) signals are increasingly used for body posture recognition.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in analyzing sEMG data.
• Understanding human-product interactions and body comfort is crucial for product design.

Purpose of the Study:

• To develop a combined approach using sEMG and CNNs for analyzing human-product interactions and body comfort.
• To compare the performance of different CNN architectures for sEMG-based posture recognition.
• To identify an optimal CNN model for accurate and efficient analysis.

Main Methods:

• A prototype usability experiment was conducted to collect sEMG measurements.
• sEMG data were categorized and split into training and testing datasets.
• Four CNN models (LeNet-5, VGGNet-11, InceptionNet V4, DenseNet) were employed for sEMG image recognition.

Main Results:

• DenseNet, a deep CNN, exhibited superior performance.
• DenseNet achieved higher accuracy with fewer layers compared to InceptionNet V4.
• DenseNet demonstrated enhanced feature reuse, easier training, regularization effects, and mitigated gradient issues.

Conclusions:

• The findings suggest DenseNet as a suitable CNN model for sEMG analysis.
• This approach can serve as a tool for assessing body comfort from sEMG signals.
• The developed models can aid in designing better human-contact products without frequent retraining.