A Study of Personal Recognition Method Based on EMG Signal.

Area of Science:

• Biometrics and Human-Computer Interaction
• Signal Processing and Machine Learning

Background:

• Traditional personal recognition methods (PINs, ID tags) are vulnerable to security breaches.
• Existing biometrics like face and fingerprint can be forged; living body features offer superior security.
• Electromyography (EMG) signals present a promising, yet under-explored, biometric for aliveness detection and spoofing prevention.

Purpose of the Study:

• To investigate and develop robust personal identification and verification methods using electromyography (EMG) signals.
• To evaluate the efficacy of different signal processing and machine learning techniques for EMG-based biometrics.
• To address the challenge of model updates in EMG-based recognition systems using transfer learning.

Main Methods:

• Surface EMG signals were collected from 21 subjects using a Myo armband under a hand-open gesture.
• Two identification methods were proposed: Discrete Wavelet Transform (DWT) with ExtraTreesClassifier and Continuous Wavelet Transform (CWT) with Convolutional Neural Networks (CNN).
• Transfer learning was applied to the CWT-CNN model for efficient model updates; a CWT-based Siamese network was developed for verification.

Main Results:

• EMG-based personal identification achieved high accuracies of 99.206% (DWT-ExtraTrees) and 99.203% (CWT-CNN).
• The CWT-CNN model, enhanced with transfer learning, effectively handled new data additions.
• The proposed EMG-based personal verification method using CWT and Siamese networks reached an accuracy of 99.285%.

Conclusions:

• EMG signals provide a highly accurate and secure biometric for personal identification and verification.
• Advanced signal processing (CWT) and deep learning (CNN, Siamese networks) are effective for EMG-based biometric systems.
• Transfer learning offers a practical solution for updating EMG recognition models, enhancing system adaptability and security.