Emotion Recognition based on Double Filtration Signals Learning Network on Different Hemispheres

Recently, emotion recognition has been diffusely applied in many fields and achieved excellent results, which is a promising research direction. However, existing researches fail to combine brain structure commendably to study brain activity signals. In this article, to capture human brain activity in the different hemispheres, we construct a double filtration signals learning network on different hemispheres (DFSLN-DH) to explore the influence of different cerebral hemispheres on subject independent emotion recognition. The proposed DFSLNDH model can not only extract the effective spatiotemporal information of Electroencephalogram (EEG) signals but also explores the effects of EEG signals in the left and right hemispheres on emotion recognition. Specifically, the EEG signals are firstly divided into three groups: left hemisphere signals, middle signals, and right hemisphere signals. Then, the double filtration signals learning module (DFSLM) is designed to mine the time-series features of each group, which could doubly screen the EEG signal to enrich the EEG feature information. Next, spatiotemporal information is extracted from the EEG features through depthwise separable convolutions. Finally, the significant emotional features learned from different groups are interacted through the fusion operator. Extensive experiments are performed on emotion datasets (SEED and SEED-IV). The experimental results suggest that the DFSLN-DH model is more competitive and can achieve better recognition performance.