Figure 1.

Schematic of the proposed CNN classification. (A) sEMG is segmented and spectrogram of each segment is calculated and normalized. Then principal component analysis (PCA) is performed to reduce the dimensionality of the spectrograms before passing them into the CNN classifier. The CNN model contains one convolutional layer (Conv Layer), two full connection layers (FC Layer) with dropout and a softmax loss layer. The network is trained using backpropagation in conjunction with the gradient descent method. (B) PCs of sEMG spectrogram are reshaped into a 2D matrix and rearranged in a way such that the most significant PC sits at the center of the matrix while the least significant PCs sit at the corner. The numbers indicate the ranking of the PCs. (C) Illustration of the convolutional layer. A 4 × 4 filter is convolved with the 5 × 5 realigned matrix, and gives a resultant 2 × 2 matrix. (D) Dropout method. In each training echo, 50% of the neurons in each layer will be randomly picked as dropout neurons and these neurons are ignored in the error propagation and weight update procedures (presented with dashed line).