Light-weight residual convolution-based capsule network for EEG emotion recognition

In recent years, electroencephalography (EEG) emotion recognition has achieved excellent progress. However, the applied shallow convolutional neural networks (CNNs) cannot characterize the spatial relations between different features well, which affects the performance of these models. In addition, because the amount of EEG sample data is small, it is challenging to collect and annotate enough EEG signals for emotion recognition. Extracting more distinguishing features from small sample data is one of the problems faced by EEG emotion recognition. To solve these problems, this paper proposes a light-weight residual convolution-based capsule network (LResCapsule) for EEG emotion recognition. The LResCapsule consists of a Light-ResNet based feature extractor and a capsule-based classifier. Because of the low EEG training data, we propose a low-parameter Light-ResNet to automatically extract deep emotion features from the raw EEG signal. Then the Capsule-based classifier is applied to identify the positional relations between local features and global features in the spatial domain, which can further improve the performance of EEG emotion recognition. Compared with ResNet18, the number of parameters of our proposed Light-ResNet is reduced by 84.5%. The experimental results on the DEAP and DREAMER datasets show that the proposed LResCapsule can outperform state-of-the-art methods in both subject-dependent and subject-independent experiments.