Research On Extracting Emotional Features From ECG By Using Wavelet Transform

Affective computing is computing in which the computer measures the user's emotional state and then uses this information in order to change the way it reacts to the user. Emotion recognition is the fundamental technique of affective computing, and the accuracy of emotional recognition depends heavily on the selection of emotional features. In the physiological signal-based emotional recognition, the key issue is to identify and extract the features from physiological signals under different emotion modes.In this master study, the ECG (Electrocardiogram) signals are used for emotion recognition and wavelet transforms are adopted to extract emotion features from the ECG under different emotion modes (pleasure, sadness etc.) in terms of wavelet efficiencies. The relationship between the value of ECG and emotion states are analyzed. The main contributions of this master study are summarized as follows:1. We apply wavelet transforms to the ECG signals that are collected from one person under different emotion states at Augsburg university of Germany. The statistic features of the wavelet coefficients are analyzed. The magnitudes of these feature coefficients under four different emotion states collected at the same day are compared and it is found that these features and their corresponding magnitudes can be used for emotion recognition. In order to improve the recognition performance, these coefficients are first normalized, and a threshold method is used to classify the emotions. The recognition over pleasure and sadness gives a better performance with an up to 92% successful recognition rate. The experiment results validate the wavelet transform-based emotion feature extraction.2. Furthermore, different wavelet decompositions are applied to the experiment data, and an adaptive value is determined by both the within-class scatter and between-class scatter. We then propose new criteria and corresponding methods to select better wavelet functions for ECG-feature extraction.3. More experiments are carried out at our lab where pleasure and sadness modes are induced by video clips and six physiological signals (including ECG) are collected through Biopac MP150 physiological signal recorder. We take use 244 sets of the collected ECG to validate our algorithms. These data sets are categorized into three groups according to the video's induce strength and the impacts of induce strength are analyzed. The experimental results show that, on the data over average induce strength, the successful recognition rate for pleasure and sadness is 71%; on the data at strong induce strength, the successful recognition rate for pleasure and sadness can reach 76%. It is shown that a better cognition can be achieved by using data with stronger induce strength.