Channel-distribution Hybrid Deep Learning for sEMG-based Gesture Recognition

In recent years, CNNs (Convolutional Neural Networks) with their powerful feature representation and feature learning capabilities, have played an important role in gesture recognition tasks based on sparse multichannel surface EMG signals. As each muscle group in the upper limb plays a different role in a particular hand movement, we propose a hybrid CNN model that considers the spatial distribution of muscle groups in the myoelectric channel to improve the accuracy of hand gesture recognition. The model takes the spectrogram of CWT (Continuous Wavelet Transform) as input, based on the spatial distribution of channels, decomposes all channels into multiple input streams, lets the CNN learn the features of each stream separately, and gradually fuses (slowly fusion) the features learned by each stream, and then performs gesture classification. Finally, the results of several of these stream-division methods are fused for decision making to obtain classification accuracies. The proposed model was validated and tested on the Nina Pro DB4 dataset, and the average accuracy was improved compared to both traditional machine learning methods and multi-stream CNN models that do not take into account the spatial distribution of channels.