【作者】 张慧玲；

【导师】 刘光远；

【作者基本信息】 西南大学 ， 信号与信息处理， 2011， 硕士

【摘要】 情感识别不仅是人工智能的基础研究领域,也是人机交互领域的一个热门话题。基于生理信号的情感识别比基于面部表情和基于语音的情感识别在实现上更为困难和复杂,但是因为生理信号由人体白发产生,不受人的主观意识的控制,因此在研究过程中更加准确和可靠。实验室建立了规模较大的情感生理信号数据库,主要研究高兴、惊奇、厌恶、悲伤、愤怒和恐惧6种情感状态下由生理过程自发产生的7种生理信号,其中包括电生理信号如心电(ECG)、脑电(EEG)、肌电(EMG)、皮肤电导(GSR)和非电生理信号如心率(Heart Rate)、脉搏(Pulse)、呼吸(RSP)。脉搏信号一直以来都是一个重要的研究课题,但是基于脉搏信号的情感识别的先行研究还较少,论文主要以脉搏信号,辅以心电信号(Electrocardiography, ECG)为研究对象,通过生理信号处理方法,对这些信号进行滤波和重构,并提取重要特征,同时将局部搜索和变异策略引入到最大最小蚂蚁系统(Max-Min Ant System, MMAS)中,结合Fisher分类器对情感进行识别,获得了较好的效果。本文主要包括了以下工作：1.对信号进行预处理：利用小波包变换去除信号的基线漂移和工频干扰,用五点三次法对信号进行平滑。2.通过小波变换准确定位脉搏波的主波波峰,用高斯函数钟形波拟合脉搏波的主波、潮波和重搏波,进而提取特征。3.因为提取的特征数较多,因此我们使用智能算法进行特征选择。为了改善最大最小蚂蚁系统收敛速度慢,易陷入局部最优解的缺点,将伪随机比例、局部搜索和变异策略引入最大最小蚂蚁系统中,得到改进的最大最小蚂蚁系统(IMMAS)。4.首先使用104个统计特征作为原始特征集,采用IMMAS算法与Fisher分类器相结合的方法,分别对情感进行“一对一”和“一对多”识别；继而将IMMAS选出的最优统计特征子集与新特征相结合,重新对高兴、惊奇、厌恶、悲伤、愤怒、恐惧进行识别,获得了更好的识别效果。实验和仿真结果表明,使用脉搏信号进行情感识别是可行的。更多还原

【Abstract】 Emotion Recognition is not only the basis of artificial intelligence, but also a hot research issue in human-computer interaction (HCI). Emotion recognition based on physiological signals is more difficult and complicated to achieve than that based on facial expressions and speech; however, because physiological signals are produced by human body spontaneously, without conscious control of the subject, so the research process is more accurate and reliable. A fairly large physiological signal database is created by our lab and 7 kinds of physiological signals, including electrophysiological signals such as electrocardiogram (ECG), electroencephalogram (EEG), electroneuromyography (EMG), Galvanic Skin Response (GSR) and non-electrophysiological signals such as heart rate, pulse signal, respiration (RSP), under 6 emotion states, namely, happiness, surprise, disgust, grief, anger and fear.For quite a long time, pulse signal has been an important research issue; however, there are few prior research of emotion recognition based on pulse signal. Pulse signal aided with ECG signal are studied in this paper, and important features are extracted from the signals after filtering and reconstruction. The Improved Max-Min Ant System (IMMAS), combined with Fisher classifier is used for emotion recognition, and good effect is obtained.The following work is discussed in this paper:1. Signal preprocessing:wavelet packet transform is used to remove baseline shift and power-line interference of the signal and five-spot triple method is adopted for signal smoothing. 2. Percussion wave crest of pulse signal is positioned accurately through wavelet transform, and then three Gaussian functions are used to fit the percussion wave, tidal wave and dicrotic wave for feature extraction.3. Because there are a large number of original features, so intelligent algorithm is used for feature selection. In order to overcome the disadvantages of MMAS, such as slow rate of convergence and liable to trap in local optimum, Pseudo-random proportion, local search and variance strategy are introduced to MMAS and the improved MMAS (IMMAS) is obtained.4. Firstly, the 104 statistical features are used as the original feature subset, and IMMAS combined with Fisher classifier is adopted one-vs-one emotion recognition and one-vs-rest emotion recognition; then the optimal feature subsets selected by IMMAS combined with some new features are used to recognize happiness, surprise, disgust, grief, anger and fear once again, and better recognition rates are obtained.The experiment and simulation results show that it is feasible to use pulse signal for feature selection.更多还原