Research On Feature Fusion For Emotion Recognition Based On Discriminative Canonical Correlation Analysis

With the rapid development of emotion recognition technology,how to achieve human-machine interaction naturalization and intellectualization,enabling human emotion state to be recognized effectively by machine,and getting natural and harmonious emotional feedback results,has become the focus of research in the field of emotion recognition.Recognition technology based on single modal has become more and more mature,but due to the acquisition of noise and user's deliberate camouflage,the recognition technology based on voice,facial expression and other signals,the results can't meet the requirements.However,the emotional recognition using physiological signals overcomes the shortcomings of speech and facial expressions,and the use of wearable devices to facilitate the acquisition will not be affected by external environment and subjective factors.Therefore,this paper presents characteristics of multimodal fusion algorithm based on physiological signals to improve the accuracy of recognition.Using two algorithms,this paper coalesces the physiological multi-modal signal,one is the canonical correlation analysis(CCA)feature information fusion method,projects two modal vector fusion,and fully considers the correlation of the two samples,rather than simply stitching and overlay;the other is DCCA,on the basis of CCA,it introduces label information and gives full consideration to the correlation within and between class of peripheral physiological signals and EEG signals,making class reaches the maximum value,minimum value between classes.Using K nearest neighbor algorithm(KNN),support vector machine classifier(SVM)and random forest classifier(RF)for the identification and classification of the fused features,and then research the effect of two fusion methods on the emotional state classification.This paper uses DEAP data set.The data set by the music video collected a total of 32 subjects of EEG and physiological signals,by using the method of integration of two kinds of features,each kind of emotion is mapped to the space dimension,analysis of the characteristics of information fusion,Arousal,valence,Dominance,liking in four dimensions,using three kinds of classification methods:K nearest neighbor algorithm(KNN),support vector machine classifier(SVM)and random forest classifier(RF)to classify emotion,thereby calculating accuracy and F1-score.The experimental results show the effect of the experiment based on the CCA feature fusion algorithm is remarkable.On the four dimensions,three classifiers are used for emotion recognition.The experimental results obtained have obvious advantages,and the accuracy is increased by nearly 8%compared with the traditional ones.The maximum accuracy for the four dimensions is 59.6%,62.2%,64.2%,and 66.9%,and the results of the F1 coefficients are 0.72,0.725,0.748,0.788.The experimental results based on DCCA feature fusion algorithm outperform those based on CCA,and the accuracy is nearly 2%higher than that of CCA feature fusion.The accuracy of Arousal,Valence,Dominance and liking dimensions is 59.2,62.6%,65.1%and 68.1%,and the calculation results for F1 coefficients are 0.735,0.736,0.758 and 0.758.