Abstract
Gesture recognition based on surface electromyographic signals (sEMG) is a widely used method for human-computer interaction, yet traditional manual feature extraction often fails to capture high-level signal complexities. This paper proposes SEDCNN-SVM, a hybrid model combining an improved Deep Convolutional Neural Network (DCNN) with a Support Vector Machine (SVM). The DCNN architecture incorporates Squeeze-and-Excitation Networks (SE-Net) and residual modules to enhance feature representation and mitigate gradient vanishing. By replacing the standard Softmax layer with an SVM, the model improves generalization and classification accuracy. Experimental results demonstrate that SEDCNN-SVM achieves a recognition accuracy of 95.5%, significantly outperforming traditional methods and proving its efficacy for real-time online recognition.
Introduction
Gesture recognition via sEMG—the electrical activity produced by skeletal muscles—is a cornerstone of modern rehabilitation and robotics. However, two major hurdles remain in existing systems. First, traditional methods rely on manual feature extraction (such as mean absolute value or root mean square), which often misses complex, high-level patterns in the signal. Second, while Deep Convolutional Neural Networks (DCNNs) can automate feature extraction, they often suffer from "gradient vanishing" as they get deeper, leading to a loss of essential signal information and lower accuracy. The authors sought to create a model that could automatically learn complex features without losing information during the process, while maintaining a stable and generalizable classification performance.
Methodology
The proposed SEDCNN-SVM architecture is a multi-stage pipeline designed to maximize both feature discovery and classification robustness.
Preprocessing
Raw sEMG data is cleaned using Butterworth high-pass filters to remove low-frequency noise (below 20Hz) and trap filters to eliminate 50Hz power-line interference. This ensures the input to the neural network is high-quality muscle activity data.
Feature Extraction (SEDCNN)
The model uses two sub-blocks (DCNN-1 and DCNN-2) with a total of four convolutional modules. Two key enhancements are integrated:
- SE-Net (Squeeze and Excitation): This adds an "attention" mechanism, allowing the network to weight the importance of different muscle channels dynamically.
- Residual Modules: These use skip-connections to prevent gradient vanishing, ensuring that original signal information is preserved as the network depth increases.
Classification (SVM)
Instead of using the standard "Softmax" layer found in most neural networks, the authors replaced the final layers with a Support Vector Machine (SVM) using a Gaussian kernel. SVMs are mathematically optimized to find the "widest margin" between classes, which leads to better generalization when dealing with the high-dimensional feature vectors extracted by the DCNN.
Experimental Setup
The study utilized a custom dataset collected via a Myo armband, an 8-channel sEMG wearable with a 200Hz sampling frequency. Data was collected from 10 healthy individuals (5 male, 5 female, aged 20–30).
The experiment involved 10 distinct hand movements, including "fist," "OK," "praise," and various finger counts. Each movement consisted of 3,200 sampling points, with 15 repetitions per gesture to ensure a robust training set.
Key Results and Discussion
The hybrid SEDCNN-SVM model demonstrated superior performance across several metrics:
- Top-Tier Accuracy: The model reached an accuracy of 95.5%, significantly outperforming standalone DCNNs (90.4%) and traditional SVMs (60.0%).
- SOTA Comparison: The model beat several State-of-the-Art (SOTA) algorithms, including Graph Convolutional Networks (GCN) and Transformer-based models (MSMHA-VTN), which scored between 90.8% and 92.6%.
- Real-Time Performance: In online testing with a robotic manipulator, the system maintained a 95% success rate (76 correct classifications out of 80 trials), proving its viability for real-world applications.
- Robustness: The model maintained high performance (above 90%) across different individual subjects, despite variations in physiological conditions.
Conclusion
The SEDCNN-SVM model successfully bridges the gap between automated feature learning and robust classification. While the results are impressive, future work will focus on addressing specific misclassifications—such as the "up" gesture, which is prone to noise—and expanding the system to recognize dynamic gestures. Further exploration of attention mechanisms beyond SE-Net may also refine how the model processes multi-channel muscle data for even higher precision in fluid human-robot interactions.
Summary updated