Research And Application Of Emotion Recognition Methods Based On Physiological Signals

The demands of human-computer interaction,medical care,social security and other fields promote the development of emotion recognition research.Emotion recognition based on physiological signals is more objective and reliable than emotion recognition based on speech,image or video.In emotion recognition based on physiological signals,emotional feature of traditional machine learning is difficult to dig out the internal mechanism of physiological signals,and emotional feature of deep learning doesn't perform well when the data volume is limited.In addition,most emotion recognition algorithm based on physiological signals is user-dependent model and poor universality,it is difficult to obtain good recognition effect on user-independent model.To solve these two problems,this thesis proposes the methods of feature fusion and domain adaptation to design the emotion recognition model based on physiological signals.The main work of this thesis is as follows:1.In order to solve the problems of shallow and deep emotional features mentioned above,this thesis fuses shallow and deep emotional features to obtain features with higher emotional representativeness.The experiment result shows that,in DEAP dataset,using the proposed fusion feature to emotion recognition increased 15.10% and 0.94% average recognition accuracy than using shallow or deep emotional features respectively;and in MAHNOB-HCI dataset,using the proposed fusion feature to emotion recognition increased 34.01% and 6.89%.It shows that the fusion emotional features effectively improve the representation of emotion.Compared with Ningjie Liu's literature in 2018,the algorithm proposed in this thesis increased by 6.38% and 5.09% respectively in arousal and valence dimensions.2.To solve the problem that the same classifier may have different recognition performance in different datasets,this thesis proposed an adaptive weighted voting decision fusion algorithm which uses adaptive weight calculation of weighted voting strategy.By fusing the labels from logistic regression,linear discriminant analysis and AdaBoost classifier,let the proposed emotion model achieve better robustness.The experiment result shows that,on the DEAP dataset,using the proposed decision fusion improved the recognition accuracy by 3.52% on average compared with using the three classifiers alone;on the MAHNOB-HCI dataset,improved the recognition accuracy by 7.27% on average.3.To solve the problem of poor universality of the emotion recognition model caused by individual differences,this thesis uses JDA algorithm for cosine similarity weighting based on adaptive threshold selection to carry out the domain adaptation between the source domain and the target domain.The parts of the source domain data with low similarity to the target domain are screened out by adaptive threshold selection,and then the joint distribution adaptive algorithm with cosine similarity weighting is used to adapt the domain between the source domain and the target domain.The experiment result shows that,on the SEED dataset,compared with baseline data before domain adaptation,the domain adaptation method using cosine similarity for weighting and screening improved the recognition accuracy by 18.80% on average.Compared with the existing algorithms on some user independent models,the recognition accuracy is improved by about 5%.