| Summary: | 碩士 === 國立成功大學 === 電腦與通信工程研究所 === 98 === For building a subject independent emotional recognition system, in this thesis, we classified four emotions and attempt to find the relevant features which can distinguish human’s emotion accurately. Different to other researches, we further classify three levels of each emotion and the physiological signal we adopt is only the ECG, due to the recent development of the ECG, it is much easier to measure and carry.
The reaction of physiological to emotions varies from person to person, the level of the emotion is hard to be consistent either. In order to find the combinations of the features which are the most distinguishable for classifying emotions, we first calculated features from frequency domain, time domain and nonlinear method in ECG. Then we applied statistical methods to calculate the value of each emotion segments and followed by z-score method for feature normalization. After features are normalized, we use feature selection combining pLDA for classification. In classification, most common classifiers are suited in handling the two-class classification. In order to raise the recognition rate, we turn the multiclass problem into a two-class issue. There are C!?2 combinations in C classes, the major problem needed to be addressed is how to decide the specific combination that can achieve the highest recognition rate. If we first separate the most distinguishable class as a group and the rest class as a group, the new data can be classified more accurately due to these two groups are much separate. Based on this consideration, by applying high confidence hierarchical extraction policy, we raise nearly 9% from the inside-tests that is 77.65% to 88.583%, and the outside-test is 79.349%. For three levels of each emotion classification, the emotion level of calm, joy, sad and fear in outside-test is 76.19%, 69.6%, 73.214% and 73.809%, respectively.
|
|---|
