【作者】 欧阳琰；

【导师】 桑农；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2013， 博士

【摘要】 人脸表情识别技术能够使计算机识别人的表情,从而营造真正和谐的人机环境。表情识别对建立友好的人机交互界面有着非同一般的重要意义。如今表情识别技术已经深入应用到了我们日常生活中的多个领域：远程教育系统、疲劳驾驶检测系统和微笑检测技术等。因此,本文着重研究了表情识别中的几项关键技术。表情识别技术不同于人脸识别和纹理识别,它拥有自己独有的定义和特性。如今,表情识别关键技术的研究重点主要体现在两个方面：(1)利用表情的独有特性(面部运动单元)来改进经典的人脸识别或者纹理识别方法；(2)模拟生物视觉系统提出表情识别方法,使该方法拥有生物视觉系统的特点,即对噪声和遮挡等具有一定的鲁棒性。本文通过研究以上两个方面中的现有先进关键技术,综合运用了数字图像处理、生物视觉感知等人工智能技术对这些关键技术进行了改进。基于面部运动单元的表情识别方法是当前表情识别方法中的关键技术之一,本文在认真分析经典算法的基础上,提出了新的特征组合策略和采用了分类能力更强的分类器训练方法,从而进一步提高了识别准确率。但是,基于面部运动单元的表情识别方法是没有考虑表情识别中的噪声和遮挡等关键问题的。因此,本文对另一项表情识别方法中的关键技术也进行了研究,即对遮挡鲁棒的表情识别技术。由于生物视觉系统能够非常轻易的分辨出带有噪声和遮挡的人脸表情,因此在模拟生物视觉系统的基础上,基于稀疏表达的分类方法被越来越多的应用到表情识别之中。本文在研究稀疏表达的基础上,根据不同的应用情况提出了多种改进思路：(1)为了提高基于稀疏表达分类方法的识别率,提出了采用方向梯度直方图描述子替代传统的特征；(2)提出了一种模拟生物视觉的表情识别模型,并以此模型为标准确定了局部二值化模式与方向梯度直方图是最佳的特征提取方式,并使用基于贝叶斯理论的分类器融合方法对基于两种特征的分类方法进行了决策级上的融合,进一步提升了识别准确率；(3)针对前面两种方法运算时间过高的问题,提出了2项特征选取准则,并根据该准则选取出了新的特征,虽然使用该特征分类方法的识别准确率不如前面两种方法,但是处理单幅图像的运算时间大幅下降,并且其识别准确率是高于现有基于稀疏表达的分类方法。(4)为了进一步提高分类方法的鲁棒性,使用了一种基于参数估计的稀疏编码求解方法,并证明该方法能够提升基于稀疏表达表情识别方法的鲁棒性。更多还原

【Abstract】 Facial expression recognition technology enables the computer to recognize the hunman facial expressions and create a truly harmonious human-machine environment. Expression recognition has extraordinary significance to establish friendly man-machine interface. Nowadays, expression recognition technology has in-depth applied to many areas of our daily lives:the distance education system, driver fatigue detection system and smile detection technology. Therefore, this paper focuses on several key technologies in facial expression recognition.Expression recognition technology is different from face recognition and texture recognition; it has its own unique definitions and characteristics. Today, the expression recognition of key technologies research focus on two aspects:(1) Use the unique characteristics of facial expressions (facial action unit) to improve the classic face recognition or texture recognition method;(2) Propose facial expression recognition method by simulating the biological visual system, the method has the characteristics of the biological visual system and has a certain robustness to noise and occlusion. We improve these tecnologies by using some artificial intelligence techniques such as digital image processing and biological visual perception, based on the research of the existing advantages of key technologies,Expression recognition method based on facial action unit is one of the key technologies in the current expression recognition method. Based on the analysis of the classical algorithm, we propsed a new feature combination strategy and a final classifying method with better classification capabilities to improve the accuracy rates of expression recognition. However, the expression recognition method based on facial action unit is not considered the corruption and occlusion problems. So, we study the other key technology in facial expression recognition which can robust to occlusion.As the biological visual system is able to very easily distinguish the facial expressions with noise and occlusion, so sparse representation based classifiers (SRC) are more and more applied to facial expression recognition.Based on the research of SRC, we proposed a variety of improvement ideas:(1) Propose to use histogram of gradient descriptor to take place of traditional feature extraction method, for the purpose of increasing the accuracy rates of SRC.(2)Propose an expression recognition model by simulating biological visual based on the existing research result. Determine that the Local Binary Patterns and histogram of gradient descriptor are the best features. For the purpose of further increasing the accuracy rates, use classifier combination method which based on Bayesian theory to fuse the results of two classifier methods.(3) Propose two feature selection criterias to solve the problem of time-consuming. Select a new feature based on these two criterias, the facial expression recognition method which based on the feature and SRC can decrease the time-consuming and give better performance than the exisiting method based on SRC.(4) For the purpose of increasing the robustness of SRC, appling a robust sparse coding model to facial expression recognition. The results show that the robustness of SRC can be improved.更多还原