A deep learning approach using attention mechanism and transfer learning for electromyographic hand gesture estimation

Accurate surface electromyography decoding of hand gestures is pivotal for advancing human-computer interaction applications. Recent developments in end-to-end deep neural networks have facilitated the automatic extraction of features from raw signals; however, not all extracted features are relevant and unimportant information can impede gesture estimation. Additionally, the accuracy of hand gesture estimation often declines when applying pre-trained models to new users. Therefore, a deep learning model using attention mechanism and transfer learning for electromyographic hand gesture estimation is proposed in this study. The proposed model consists of a feature extractor, a label classifier and a gesture estimator. A convolutional neural network is employed as the feature extractor, enabling the learning of high-level discriminative features from input signals. Attention modules are integrated into the convolutional layers, enhancing the extraction of pertinent features by emphasizing critical information through weight recalibration. The label classifier, consisting of three fully connected layers, classifies gesture labels using feature maps generated by the feature extractor. The gesture estimator, which adopts a threshold voting algorithm, is capable of estimating gestures during hand movement, making it suitable for real-time hand gesture estimation. A transfer learning approach is utilized to selectively transfer parameters from a pre-trained model to a target model, which is then partially retrained with limited target data. The proposed model is evaluated on both the Myo dataset and the public NinaPro dataset, demonstrating superior performance compared to baseline models in terms of estimation accuracy. Furthermore, the proposed transfer learning method proves more effective than training a new model from scratch, as retraining the target model with only two sessions yields satisfactory estimation accuracy.