Multi-Scale Entropy Algorithm And Its Application In Emotion Recognition

The further development of human-computer interaction and artificial intelligence,the research of emotional computing is becoming more and more attention. Emotion recognition as the main content of affective computing, including the identification of facial expressions, body posture, voice, and other aspects of physiological signals. In the study of emotion recognition based on physiological signals, the common used signals,such as ECG, EMG, EEG, GSR, breathing. EEG is a comprehensive reflection of the electrical activity of brain neurons, it can reflect the brain respond in different cognitive tasks and functional status. And the EEG signals not like the facial expression, body posture, which can be deliberately hide, so the research of emotion recognition based on EEG, is more objectively and scientifically.The main features of EEG signals are nonlinear, multi-scale and multi-resolution, and the multiscale entropy of EEG can reflect the characteristics of EEG fully. In this paper,the recognition of three emotional states of positive, neutral and negative is based on multi scale entropy of EEG. There are totally of 12 methods of multiscale entropy, each of them is consists of the one method from coarse-grained method, moving-average method,discrete wavelet transform, empirical mode decomposition, and the one from approximate entropy, sample entropy, permutation entropy. Compare the performance of different multiscale entropy in emotion recognition, based on the accuracy of classification and the time complexity of the algorithm, four best methods of multiscale entropy are obtained,CG-PE, MA-PE, CG-Ap En, CG-Samp En. The length of sequence which is processed by multiscale method will greatly reduce, and result in the information loss, according to this issuse, we use the adaptive binarization method to capture the small changes of sequence.Contrast the methods and improved methods, the classification accuracy of improved methods is increase by 16.78% at most, this result show that the adaptive binarization can indeed caught the small changes, especially in the coarse-grained process. In order to explore the role of EEG at each electrode, compare the ability of 16 electrodes, found thatO2 have the highest classification accuracy, up to 75.24%. The occipital region is the most sensitive to the change of emotion in all the brain region.