Deep end-to-end transfer learning for robust inter-subject and inter-day hand gesture recognition using surface EMG

Jianfeng Li,Xinyu Jiang,Jiahao Fan,Yanjuan Geng,Fumin Jia,Chenyun Dai

DOI: https://doi.org/10.1016/j.bspc.2024.106892

IF: 5.1

2024-10-02

Biomedical Signal Processing and Control

Abstract:With the increasing demands on human–machine interaction in intelligent neural rehabilitation, surface electromyogram (sEMG) decoders enabling hand gesture recognition have attracted more attentions. Although previous studies have proposed diverse models, their low inter-subject and inter-day robustness still cannot satisfy the demands in real world applications. In this work, we employed deep transfer learning to adapt the mismatch between source and target domains for robust hand gesture recognition. We systematically investigated 5 deep transfer learning algorithms on end-to-end convolutional neural network (CNN) architecture. Statistical machine learning methods with benchmark handcrafted features and general domain adaptation algorithms were also implemented as the strong baseline performance for the comparisons. The performance of the model under both inter-subject and inter-day applications was evaluated using high-density sEMG (HDsEMG) signals with 256 channels acquired from 36 subjects twice on 2 days in various rigorous validation protocols. Importantly, we examined why the deep transfer learning functioned by monitoring both the intra-domain and inter-domain mismatch between the learned features during model training. The results demonstrated that, without deep transfer learning, the intra-domain mismatch between samples under the same gestures decreased but the mismatch between domains (source and target domains) increased rapidly. The clustering of samples under the same gesture came at the expense of higher mismatch between different domains. With deep transfer learning applied, the inter-day and inter-subject performance of CNN was greatly improved (improved from 85.4% to 91.39%, and from 82.22% to 91.38%, respectively), which also outperformed benchmark statistical machine learning models with domain adaptation. In addition, the network structure of the proposed method is simple and can be easily implemented online. Overall, the findings in this work show great promise of transfer learning in hand gesture recognition under inter-subject and inter-day settings, significantly reducing the inconvenience of model re-calibration in practical use.

engineering, biomedical