Brain functional connectivity analysis and crucial channel
selection, play an important role in brain working principle exploration
and EEG-based emotion recognition. Towards this purpose, a novel
channel-wise convolution neural network (CWCNN) is proposed, where
every group convolution operator is imposed only on a separate channel. The inputs and weights of the full connection layer are visualized
by using the brain topographic maps to analyze brain functional connectivity and select the crucial channels. Experiments are carried out
on the SJTU emotion EEG database (SEED). The results demonstrate
that positive and neutral emotions evoke greater brain activities than
negative emotions in the left frontal region, which is consistent with the
result from the power spectrum analysis in the literature. Meanwhile, 16
crucial channels, which are mainly distributed in the frontal and temporal regions, are selected based on the proposed method to improve
emotion recognition performance. The classification accuracy by using
the selected crucial channels is similar to that without channel selection.
But the model with the 16 selected channels is more memory-efficient
and the computation time can be reduced substantially.