Resumen
Electroencephalography (EEG)-based emotion recognition technologies can effectively help robots to perceive human behavior, which have attracted extensive attention in human?machine interaction (HMI). Due to the complexity of EEG data, current researchers tend to extract different types of hand-crafted features and connect all frequency bands for further study. However, this may result in the loss of some discriminative information of frequency band combinations and make the classification models unable to obtain the best results. In order to recognize emotions accurately, this paper designs a novel EEG-based emotion recognition framework using complementary information of frequency bands. First, after the features of the preprocessed EEG data are extracted, the combinations of all the adjacent frequency bands in different scales are obtained through permutation and reorganization. Subsequently, the improved classification method, homogeneous-collaboration-representation-based classification, is used to obtain the classification results of each combination. Finally, the circular multi-grained ensemble learning method is put forward to re-exact the characteristics of each result and merge the machine learning methods and simple majority voting for the decision fusion. In the experiment, the classification accuracies of our framework in arousal and valence on the DEAP database are 95.09% and 94.38% respectively, and that in the four classification problems on the SEED IV database is 96.37%.