Improve inter-day hand gesture recognition via convolutional neural network based feature fusion

The learning of inter-day representation of electromyographic (EMG) signals across multiple days remain a challenging topic and not fully accommodated yet. This study aims to improve the inter-day hand motion classification accuracy via convolutional neural network (CNN) based data feature fusion. An EMG database (ISRMyo-I) were recorded from six subjects on ten days via a low density electrode setting. This study investigated CNNs’ capability of feature learning, and found that the output of the first fully connected layer (CNNFeats) was a decent supplement feature set to the most prevalent Hudgins’ time domain features in combination with 4th order autoregressive coefficients (TDAR). Through adding the automatically learned CNNFeats to the handcrafted TDAR feature set, both linear discriminant analysis (LDA) and support vector machine (SVM) classifiers received >3% accuracy improvement. Similarly, taking TDAR as additional input to the CNN improved the accuracy by >1% in the comparison with the basic CNN. Our results also demonstrated that the CNN approach outperformed conventional approaches when multiple subjects’ data were available for training, while traditionalapproaches were more adept at presenting motion patterns for single subject. A preliminary conclusion is drawn that substantial ‘common knowledge/features’ can be learned by CNNs from the raw EMG signals across multiple days and multiple subjects, and thus it is believed that a pre-trained CNN model would contribute to higher accuracy as well as the reduction of learning burden