Surface Electromyography Relation Network

Metric-based learning combined with few-shots learning is efficient in learning similarity. This paper presents a metric learning based methodology for hand gesture recognition based on surface electromyography (sEMG) signals. We designed a model based on the relation network architecture to generate relation scores for assessing the similarity between the input sEMG signals, thereby achieving hand gesture recognition. In this proposed Surface Electromyography Relation Network (SE-RN), we first extracted the features of different input signals through a feature extraction part. Then, we concatenated this extracted information with the features from the referenced sEMG signals before feeding it into the relation module. Finally, with the relation module, we determined the similarity levels through the relation scores generated by the final classification layer. We performed experiments on C (C=5, 10) way one-shot experiment, and results showed that the designed SE-RN can achieve accuracies of 0.922~0.931 in these two experiments.