RGB-Style Input Representations for EEG: Evaluating Spatial Concatenation Versus Band-Wise Stacking in Deep Emotion Recognition.

BACKGROUND/

objectivesElectroencephalography (EEG) is widely applied in emotion recognition. Integrating diverse frequency and spatial features to improve performance remains a major challenge.

methodsThis paper proposes two preprocessing methods to map EEG signals into image-style representations. These methods preserve the spatial topology and enable effective feature extraction using convolutional neural networks. The first method is a spatial concatenation method (SCM). It projects three feature types onto color channels, providing a structural prior that encourages the network to learn the three feature types within local spatial windows. It differs from traditional spectral mixing, which maps frequency bands to color channels. The second method is a band-wise stacking method (BSM). It treats frequency bands as independent depth frames to form a three-dimensional tensor. This structure is designed to facilitate the learning of inter-band relationships while preserving band-specific information. Dedicated convolutional neural network architectures are designed for these tensor structures, aligned with the spatial and spectral organization of the proposed SCM and BSM.

resultsExperiments on the DEAP and DREAMER datasets for binary Arousal and Valence classification show that both representations achieve competitive results. The BSM achieves higher accuracy than the SCM on the DREAMER dataset, while both methods perform comparably on the DEAP dataset.

conclusionsThe proposed strategies offer efficient convolutional neural network approaches for EEG emotion recognition systems.