Studies of Affective Recognition from Pulse Signal Based on Correlation Analysis and Ant Colony Optimization Algorithm
|Course||Signal and Information Processing|
|Keywords||Affective Recognition Pulse Signal Feature Selection Correlation Analysis Max-min Ant Colony Algorithm|
Nowadays, with the rapid development of science and technology, the human-computer interaction field has been no longer strange. As a key of human-computer interaction, affective computing has been widely applied to many fields, and great importance has been attached to affective recognition which is one of the important parts in the field of affective computing. At present, the affective recognition is mainly achieved through the collection and analysis of image signals, speech signals, text signals and physiological signals. These ways not only have their own strengths, but also complement each other, that is because there have been no method which can independently and accurately determine the affective characteristics. As we all know, image signals, speech signals and text signals can be generally obtained through the multimedia approach, and physiological signals come from the sensors through a non-invasive way, all of these signals can be analyzed and calculated by computer. Obviously, when people have emotions, physiological signals are directly produced by human body, which can more accurately judge people’s emotions and even the potential emotions that the first three methods can not be observed. Pulse signal is one of physiological signals, but little research has been done in the field of pulse signal-based affective recognition. The research work in this thesis is commenced by focusing on the affective recognition problem of the pulse signal.After fully research of the existing literatures, it is found that pulse is a feeble bioelectricity signal which contains a wealth of physiological and pathological information. If some features or feature combinations which represent specific emotional states can be found through pulse signal, affective recognition will be realized by getting the mapping relation between affective physiological features and emotions. The implementation process is described as follows:(1) Signal acquisition. Through setting of the appropriate experiment about pulse signal acquisition, both visual and auditory stimulation are used for arousing the subjects’emotions so that the emotion-induced pulse signals are recorded. At last,242 groups of effective affective physiological response samples of the pulse signals are obtained, which include seven emotions (calm, happiness, surprise, disgust, grief, anger and fear).(2) Feature extraction. The original pulse signals are usually subject to the interference of various factors, so preprocessing must be used in the original samples before feature extraction to ensure more reliable data sources. Then 104 statistical features are extracted from the 242 preprocessed samples, which included 20 time domain features (maximum, minimum, mean, median, standard deviation of the main wave peak, and so on)and 84 wavelet coefficient features after doing 7-layers wavelet decomposition. Furthermore, the monstrous values are deleted from these features and the features values are normalized between 0 and 1 to get original feature sets.(3) Feature Selection. A new two-stage feature selection method which combines correlation analysis with ant colony optimization algorithm is proposed in this paper. Firstly, sequential backward selection (SBS) is used for sorting of the original features. Secondly, the linear correlation coefficient is adopted to compute the correlation degrees between the features and features with high correlation degrees are removed through the result of sorting. Finally, max-min ant colony algorithm correlation analysis is used for feature selection, which searches for an optimal subset based on the compact feature subset, while the fitness is valued by both the recognition rate of Fisher classifier and the number of feature subset, and then six emotions (happiness, surprise, disgust, grief, anger and fear) are recognized by means of Fisher classifier. Furthermore, support vector machine (SVM) based on convex optimization is also applied to feature selection to compare with the result of proposed approach.(4) Model establishment. The experiments are completed on the basis of the above theory, and feature combinations which represent a specific emotional state (happiness, surprise, disgust, grief, anger and fear) can be determined through analysis. At last, two-class affective recognition model of each emotional state based on pulse signals can be established.All the experimental results show that the correlation analysis and ant colony optimization algorithm combined with the Fisher classifier can be effectively used for feature selection and classification based on affective pulse signal, and feature combinations with more stability and better performance can be obtained from the original feature sets, then the mapping relation between affective physiological features and emotions based on pulse signal can be constructed to establish effective affective recognition model. In other words, the results prove the validity of the method and a solid foundation has been established for the construction of affective recognition model.