EEG Emotion Recognition Based on 3-D Feature Representation and Dilated Fully Convolutional Networks

Abstract

Emotion recognition involving high-dimensional electroencephalogram (EEG) data demands urgently for a way to learn robust and representative EEG features for final classification. In this article, a novel framework combining 3-D feature representation and dilated fully convolutional network (3DFR-DFCN) is proposed for EEG emotion recognition (EER). To excavate the prior knowledge, such as interchannel and interfrequency-band correlation information, 1-D feature sequences are extended into 2-D electrode meshes of different frequency bands. Then, the acquired electrode meshes under multiple activation patterns are further constructed into 3-D EEG arrays to capture their complementary information. To realize cross-band and cross-channel feature learning, a dilated fully convolutional network (DFCN) is built to process the input feature array, then the spectral norm regularization (SNR) item is introduced to reduce the sensitivity to the disturbed EEG features. Both subject-dependent and subject-independent experiments have conducted on DEAP and DREAMER data sets. An average accuracy of 94.59%/81.03%, 95.32%/79.91%, 94.78%/80.23% are, respectively, obtained for valence, arousal, and dominance classifications for two kinds of experiments on the DEAP data set. The integration of spatial information and frequency-band information is meaningful for assessment of human emotional states in practical or clinical applications.