Emotion Recognition in Valence-Arousal Space from Multi-channel EEG data and Wavelet based Deep Learning Framework

The conventional emotion recognition methods are mostly based on the frequency characteristics of electroencephalograph (EEG) signals. However, spatial features are likewise valuable as it contains latent information related to emotional states. In this paper, a wavelet-based Deep Learning framework proposed by considering both frequency and spatial characteristics of multi-channel EEG signal for emotion recognition. The Continuous Wavelet Transform is utilized to produce Scalogram, a function of frequency and time to getting better time localization for short-duration, high-frequency events, and better frequency localization for low-frequency, longer-duration events. Then, the GoogleNet model is presented to recognize emotion states from Scalogram. The experiments performed with benchmark DEAP database having a three-dimensional valence, arousal, and dominance data along with multi-channel EEG data. The experimental results demonstrate that the characteristics contained in the Scalogram were complementary, and GoogleNet is more suitable for emotion recognition in two/ three-dimension space.