Enhanced Lightweight CNN Using Joint Classification with Averaging Probability for sEMG-Based Subject-Independent Hand Gesture Recognition

• Published in: IEEE sensors journal Vol. 23; no. 17; p. 1
• Main Authors: Leelakittisin, Benjakarn, Trakulruangroj, Manatsanan, Sangnark, Soravitt, Wilaiprasitporn, Theerawit, Sudhawiyangkul, Thapanun
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.09.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Artificial neural networks; Brain modeling; Classification; Computational modeling; Computing costs; Convolution; deep convolutional neural network; Density; Edge computing; Electrodes; Electromyography; Feature extraction; Floating point arithmetic; Gesture recognition; hand gesture recognition; High definition; Inference; joint-classification; Mathematical models; Muscles; Neural networks; Optimization; Parameters; Performance evaluation; subject-independent
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3296649

The existing deep convolutional neural network (DCNN) models used for hand gesture recognition based on surface electromyography (sEMG) require high computational costs. Moreover, there is a lack of a comprehensive DCNN model that can handle both high-definition sEMG (HD-sEMG) and low-definition sEMG (LD-sEMG) in a subject-independent manner. To address these issues, this study proposes a lightweight convolutional neural network model for sEMG-based subject-independent hand gesture recognition evaluated in HD- and LD-sEMG. In addition, we add a technique, Joint Classification with Averaging Probability (JCAP), to enhance the final recognition accuracy with less computational costs. We conducted three experiments (Exp-I to III): (I) optimization of the proposed model; (II) comparison with benchmark models; and (III) evaluation of model performance on simulated real-time scenario. For the results, our model achieved significantly better accuracy for all selected gestures, while computational complexity was considered low, measured via total parameters, inference time, FLOPs, and selection time. In Exp-II, our best-proposed model from Exp-I got the highest accuracy at ISRMyo-I, 85.75%, eight times smaller in terms of the number of parameters and reduces more than 94.8% of FLOPs, whereas inference time is around 20% faster compared to the smallest and fastest baseline method, respectively. The selection time of our best-proposed model was more than six times faster than the existing lightweight model in Exp-III. These strengths provide our model advantages in computational-resource-limited sEMG-based human-machine interface applications, such as edge computing, the future trend for consumer electronics.