【作者】 胡秀丽；

【导师】 叶吉祥；

【作者基本信息】 长沙理工大学 ， 计算机应用技术， 2008， 硕士

【摘要】 由于计算机在图像领域的优异性,情感计算越来越受到国内外研究者的关注,情感计算是试图使计算机能够像人类那样具有理解和表达情感能力的一个多学科交叉的新研究领域,在智能人机交互中起着非常重要的作用。人脸情感识别的研究在人工智能、计算机视觉、图像处理、模式识别、心理学等各种相关学科领域中都很有价值。人脸情感识别是模式识别领域中一个比较活跃的研究方向。从广义上讲静态图像的人脸情感识别包括:预处理,人脸检测,情感特征提取,情感分类。本文由于采用的日本人脸库(JAFFE)是单人脸的灰度图像,所以主要从三个环节上进行了研究工作,研究内容体现在以下几个方面:(1)首先,对人脸情感图像进行分割、消噪处理;然后对其作标准化处理,以便提取有效的特征,标准化处理包括尺度归一化和灰度均衡化。尺度归一化是将所有情感图像变换成标准尺寸的图像,并将所有图像中已标记的对应关键特征点校准到同一位置,灰度均衡化是将尺度归一化后的图像变换成同样的灰度级;最后对归一化后的图像使用固定像素的网格进一步分割情感子图像。(2)提出了一种新颖的情感特征提取方法——Gabor小波变换与分形维数相结合。对预处理后的情感子图像的每一个网格进行Gabor小波变换以及分形维数的计算,为提高算法效率和减少特征维数,取Gabor变换后的小波系数模的均值、方差和计算所得的分形维数作为该网格的情感特征向量。(3)通过采用Matlab中神经网络工具箱的BP、RBF神经网络,实现情感特征的多分类任务,因为使用Matlab可以方便地采用不同的核函数、激励函数。采用本文的预处理、特征提取和情感分类方法,可以实现快速鲁棒的识别工作。实验结果验证了方法的有效性。更多还原

【Abstract】 Affective computing is more and more noticed by many researchers, because computer has an advantage in image area. Affective computing is a multidisciplinary synthetic area, which tries to enable computer to have the ability of understanding and expressing affection, just like human beings. And it plays an important role in intelligent human computer interface. Facial affective recognition is very valuable in the interrelated areas of artificial intelligence, computer vision, image management, pattern recognition, and psychology and so on.Facial affective recognition is an active research area in pattern recognition. In broad sense, facial affective recognition of static images includes pretreatment, face detection, affective feature extraction and affective classification. In this paper, because each image only has a face and it is gray, we apply JAFFE database to making research mainly in the following aspects:(1) Firstly, we segment each face from initial images and get rid of yawp. Then make a standardization management so as to attract effective features, which include size standardization and gray standardization. Size standardization is that all images are transformed to the standard size and the marked corresponding key feature points are also adjusted to fixed position. Gray standardization is that the gray of all images proceeded above is transformed to the same gray level. Finally, we plot out each image further by fixed pixels.(2)We present a particular feature extraction algorithm—Gabor wavelet transformation combines with fractal dimension. For each segmented grid, we make Gabor wavelet transformation and fractal dimension compute. To improve efficiency and low the dimension, we regard the average and variance of the module from Gabor wavelet transformation along with the results from fractal dimension compute as affective feature vector of each grid.(3)We solve the multi-classification task by using back propagation neural networks and radial basis function neural networks in Matlab, because Matlab is convenient in applying different kernel functions and incentive functions.By using these methods of pretreatment, feature extraction algorithm and classification algorithm, the experiment can realize a fast and robust recognition task. The experiment results have testified the efficiency.更多还原