Accelerating 3D Convolutional Neural Network with Channel Bottleneck Module for EEG-Based Emotion Recognition

Figures

• 

  Figure 1. Overview of the workflow for EEG-based emotion recognition. 3D representation of EEG signals is constructed as the input of our proposed CNN-BN model. The model consists of a convolutional block, five consecutive bottleneck blocks, and a dense block for EEG-based emotion recognition. Conv, BatchNorm, MaxPool, and FC represent the convolution, batch normalization, max pooling, and fully connected layer, respectively. Cin and Cout are the input and output channels at each layer, respectively. Cbn denotes the reduced channel within a bottleneck block. Cmid is the reduced channel by the first fully connected layer. ReLU represents the rectified linear unit activation function.

• 

  Figure 2. (a) 2D representation of 32 channels according to the international 10–20 system. The names of the electrodes are annotated in the 2D map. The EEG signals at a time point from all EEG channels are transformed into a 2D map (9 × 9 matrix) according to the positions of electrodes on the brain. (b) The process of 3D map representation of EEG signals. The 3D EEG spatiotemporal representation serves as the input of the proposed CNN-BN model.

• 

  Figure 3. Illustration of a channel bottleneck block. Cin, Cbn, and Cout represent the number of input channels, reduced channels, and output channels within the channel bottleneck block, respectively. D, W, and H denote the length, width, and height of the feature map, respectively. The first 1 × 1 × 1 convolution reduces the number of channels in the input data from the preceding layer. The 7 × 3 × 3 convolution is then computed on the reduced-channel layer. Finally, the last 1 × 1 × 1 convolution increases the number of channels in the output data.

• 

  Figure 4. Valence–arousal model for emotional representation.

• 

  Figure 5. (a) 3D CNN architecture without a channel bottleneck; (b) Long Short-Term Memory (LSTM) cell architecture; (c) LSTM model architecture. Conv, BatchNorm, MaxPool, and FC represent the convolution, batch normalization, max pooling, and fully connected layer, respectively. Cin and Cout are the input and output channels at each layer, respectively. Cmid is the reduced channel by the first fully connected layer. ReLU represents the rectified linear unit activation function.

• 

  Figure 6. Confusion matrices of the binary emotion classification for (a) valence and (b) arousal for every trial (5 trials) for the LSTM model (top) and C3D model (middle) considered as baselines and the proposed CNN-BN model (bottom).