Hand Gesture Recognition Using Compact CNN Via Surface Electromyography Signals

Abstract

By training the deep neural network model, the hidden features in Surface Electromyography(sEMG) signals can be extracted. The motion intention of the human can be predicted by analysis of sEMG. However, the models recently proposed by researchers often have a large number of parameters. Therefore, we designed a compact Convolution Neural Network (CNN) model, which not only improves the classification accuracy but also reduces the number of parameters in the model. Our proposed model was validated on the Ninapro DB5 Dataset and the Myo Dataset. The classification accuracy of gesture recognition achieved good results.

Keywords: convolution neural networks (CNNs); hand gesture recognition; surface electromyography (sEMG).

Figure A1

Schematic of continuous signal recognition.

Figure 1

Surface Electromyography(sEMG) signals collection and classification process. Myo armband was used to collect the original signal, and then the collected signal was filtered and sampled to get sEMG signals. Continuous wavelet transform was selected to obtain the signal spectrum, and the neural network model was applied to classify the spectrum to achieve gesture recognition.

Figure 2

LCNN Architecture diagram, the LCNN consists of 2 LSTM layers, 2 one- dimensional convolution layers and 1 output layer. We use 2 LSTM layers, and each LSTM layer has 52 cells, and every cell has 64 hidden layers.

Figure 3

Schematic diagram of ConvNet architecture. In this figure, Conv refer to Convolution and F.C. to Fully Connected layers.

Figure 4

The 5 hand/wrist gestures and Myo armband. In this figure, the left is a schematic diagram of five gestures, and the right is the Myo armband.

Figure 5

Part (a) shows the waveform of sEMG signals and Part (b) the spectrum of the sEMG signals shown in (a) after wavelet transformation.

Figure 6

EMGNet Architecture contains four convolutional layers and a max pooling layer without using the full connection layer as the final output. In this figure, Conv refer to Convolution and avg_pool to a max pooling layer.

Figure 7

The 7 hand/wrist gestures in the Myo Dataset. In the Myo dataset, seven gestures are included: Netral, Hand Close, Wrist Extension, Ulnar Deviation, Hand Open, Wrist Flexion, and Radial Deviation.

Figure 8

The gesture categories in the Exercise A dataset.

Figure 9

The loss and accuracy curves during training and testing on the Myo Dataset. when the training reaches convergence, there is no phenomenon where the accuracy of the training set is high and the accuracy of the test set is low, which is over-fitting.

Figure 10

The loss and accuracy curves during training and testing on the exercise A of the NinaPro DB5 Dataset. During training and testing on the exercise A of the NinaPro DB5 Dataset, the phenomenon of overfitting appears.

Figure 11

The average accuracy of three subsets on the DB5 Dataset.

Figure 12

Parts (a) and (b) represent pictures of two different gestures in exercise A, Parts (c) and (d) show the spectrum diagrams of the sEMG signal generated by (a) and (b) respectively.