Study On Music Evoked Emotion Recognition Based On EEG Analysis

General music recommendation algorithm is based on similarity measurement of user preference,or according to music style,or genre classification.It can't satisfy the emotional needs of people.And even if it contains emotional labels,the emotions which people get are different,and can't be personalized recommendation.Because different people have different music perception.How to create a personalized emotion classification model is the most crucial step for recommendation.This research is based on a supervised machine learning algorithm.In addition to the traditional psychological scale,I also use EEG signal as an objective criterion for judging emotions.Firstly,I use EMD decomposition algorithm to analysis EEG signals to guarantee the accuracy of time and frequency resolution.I decomposed it according to the physiological indices and calculated sample entropy and frequency band energy ratio as features.Finally I use a traditional machine learning algorithm to establish a general emotional classification model.The accuracy of this model is close to 90%in all the three emotional dimensions of VAD models.Those classification results can be used as an emotional label which is more objective and convenient than the subjective psychological scale.It can avoid the difference of subjective emotion cognition.Then I created a personalized emotional-musical individual model by analyzing the correlation between individual emotions and corresponding music to find some important music features.In addition to general music features such as ZCR,MFCC,and Tempo,the more complex audio features such as Chord,CQT-Spectrum,etc.are used too.And I also try to use deep learning algorithm to capture more detail features besides the machine learning algorithm.By comparing the model result on public data sets and self-built data sets,i found that there is no obvious difference between deep learning and machine learning in the classification accuracy.It may limited by the small number of training samples,and the final classification accuracy is not unsatisfactory.However,we can find the audio features that affect people's emotional experience are different,and the number of these features can be reduced to less than 15 for everyone by using feature selection algorithm.Moreover,the similarity between the selected features of every subject and their similarity to the musical emotional experience also showed the same trend.It indicated that the algorithm is valuable for the personality recommendation.