Finger motion estimation based on frequency conversion of EMG signals and image recognition using convolutional neural network

We describe a method for estimating finger motion on the basis of the frequency conversion of electromyogram (EMG) signals and the image recognition by using a convolutional neural network (CNN). Since EMG signals are generated before finger motion, various EMG-based systems have been developed for smoothly controlling a robot hand. We used a simple CNN model for estimating finger motion by classifying images generated from a wavelet transform of EMG signals. The model has originally been used for document recognition, and it contains two pairs of convolution and pooling layers and two fully connected layers. A prototype system composed of inexpensive sensor devices was fabricated for acquiring EMG signals and capturing finger motion. The experimental results show that the test accuracy reached 83% in classifying EMG signals into four types; when a thumb opens or is closed, and fingers, except for the thumb, open or are closed.