The Research Of Feature Selection For Emotion Recognition Based On Physiological Signals

Video material is used to arouse six emotions:joy, surprise, disgust, grief, anger and fear in the emotion recognition experiment. Physiological samples are obtained which record the galvanic skin response (GSR), heart rate, pulse, respiration (RSP), electrocardiogram (ECG) and electromyography (EMG).In the previous work, the initial features are extracted from more than 200 undergraduate student samples. For the problem of feature selection algorithm for emotion recognition based on physiological signals, genetic algorithm based on simulated-annealing method, max-min ant colony algorithm, particle swarm algorithm are used. However, feature dimension is too high and the feature subset is not the same after each run of the algorithm. For the weaknesses of this type of feature selection algorithm, mRMR, SVM-REF,â–½SVM-REF and SVM-Weighted algorithms are used in the latter work and the initial features are obtained from more than 400 undergraduate student samples. Fisher and SVM classifiers are used to classify six emotions and a higher recognition rate is got. At last effective feature subset which can identify the emotion recognition system model with better performance has been found. The following works have been done:(1) Galvanic skin response (GSR), heart rate(HR), respiration (RSP), electrocardiography (ECG), pulse and electromyography (EMG) are pretreated:GSR signal is separated into continuous signals of tonic and phasic activity with a deconvolution approach. EMG and ECG signals are decomposed and reconstructed by wavelet transform. RSP and pulse signals are removed noise by filter. At last the frequency domain features and the time domain statistical features of several signals are extracted.(2) Because of the lack of the prior knowledge for emotional identification, a large number of original features have been extracted. Original feature-dimension is too high, so the feature selection process foe the emotion recognition is an NP hard problem. Genetic algorithm based on simulated-annealing method, max-min ant colony algorithm, particle swarm algorithm had been used in the previous work. However, due to the randomness of this type algorithm, the feature subset is not the same after each run of the algorithm. Because of a lot of iterative process, the time-consuming is too large and computational complexity is too high.(3) Refer to state-of-the-art feature selection algorithm, the mRMR, SVM-REF,â–½SVM-REF and SVM-Weighted algorithms have been used. Compared with the previous algorithm, more useful research in the evaluation standards is done. The search strategy is not overly complex, so the tradeoff has been made between the algorithm accuracy and complexity. In the practical application, good results have been achieved.