【作者】 刘帅师；

【导师】 田彦涛；

【作者基本信息】 吉林大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 表情是人类用来表达情绪的一种有效手段，是人机交互与情感计算研究的重要组成部分。人脸表情识别系统具有广阔的应用前景和实用价值，是当前国内外人工智能和模式识别领域的研究热点之一。由于人类表情的复杂性和微妙性使得表情识别变得很具有挑战性与艰巨性。就整个识别系统而言，虽然目前处理技术获得长足发展，但是从识别系统的性能来看，与实用化还有一定的距离，仍存在很多问题需要深入的研究。其中一个重要原因就是人脸表情识别问题会受到许多因素的影响，比如光照变化、脸部遮挡、姿态变化等。本文从表情识别的鲁棒性入手，主要研究在非均匀光照和局部遮挡情况下的人脸表情识别问题，意在提高表情识别正确率的同时增强算法的鲁棒性。本文主要针对静态人脸表情图片的特征提取与表情分类等问题进行了深入研究，研究内容与创新性工作主要如下：第一，由于表情变化富含纹理信息，而且不同的表情行为包含不同的尺度信息，因此采用Gabor滤波器提取人脸面部表情特征。针对传统Gabor特征表征面部表情全局能力弱并且特征数据存在一定的冗余的问题，提出了基于Gabor多方向特征融合与分块直方图相结合的表情特征提取方法。为了提取局部方向信息并降低特征维数，同时考虑到人脸表情行为的多尺度特性，提出两种特征融合规则，将同一尺度不同方向的Gabor特征进行融合，有效的降低了特征维数，减少了计算量和内存的需求。同时，直方图能够有效的表征图像全局特征，将Gabor特征与分块的直方图结合起来，可以多层次、多分辨率地表征人脸表情局部特征及局部邻域内的特征。实验结果表明它可以有效地提取人脸表情特征，提高人脸表情识别的精确度。第二，针对光照变化对表情识别带来的负面影响，以及传统的二维图像光照预处理方法会降低原始图像质量，丢失部分有效的辨识信息的缺点，提出了基于对称双线性模型的光照鲁棒性表情识别方法。通过双线性模型将带有未知光照的人脸表情图片中相互独立的光照与表情信息进行分离，分别构建独立的光照子空间与表情子空间，从而达到独立分析与处理的目的。通过将商光照的概念引入到双线性模型的框架中。使得所有未知光照的待测试表情图像转换到若干已知相同的光照平台上，令所有测试图像具有归一化的特性。同时，用转换后的多幅表情图像来表征原始的一张表情图像，这样就能将多幅表情特征统计起来，能使表情变化的有效辨识信息得到累加，增强表情图像的区分度，有效地提高分类精度。第三，针对独立个体面部存在部分遮挡物而不易进行人脸表情识别，且全局特征对于面部局部遮挡不具备鲁棒性的问题，提出了一种基于局部Gabor特征径向网格编码的局部遮挡表情特征提取方法。根据视网膜和视皮层上两个相邻的细胞的感受野存在部分重复的特性，将人脸表情图像分割成相邻区域存在50%重叠的子块，对每个子块内的Gabor特征采用径向网格编码策略。这样既能有效的模拟视网膜的成像，又能够降低Gabor特征数据的冗余。所得到的特征向量对于面部存在部分遮挡人脸表情具有很高的辨别能力。第四，由于面部存在遮挡时的表情特征提取分析采用的是局部特征，针对全局核支持向量机无法处理局部特征，容易受到噪声和遮挡的影响，对遮挡不具鲁棒性的缺点，提出基于局部累加核支持向量机的人脸表情分类方法。局部累加核满足Mercer理论，可以确保获得全局最优解。将由面部存在遮挡的表情图像获取的各个局部特征作为支持向量机局部核的输入，利用局部核来处理局部特征，最后对所有局部核输出进行累加整合，实现了对部分遮挡人脸表情的鲁棒性识别。通过实验验证，本文所提的局部径向基累加核支持向量机和局部归一化线性累加核支持向量机策略简单有效，易于应用。而且本方法不仅能够较好的识别部分遮挡人脸表情，对于无遮挡人脸表情的识别率也高于传统的全局核SVM。最后，总结了全文所做的工作，提出了今后进一步需要研究的问题。更多还原

【Abstract】 Facial expression is an effective way for humans to express their emotions, which is animportant part of human-computer interaction and affective computing. At present, facialexpression recognition system has broad application prospects and practical value, which isan active topic in the field of artificial intelligence and pattern recognition in the world. It isvery diffcult to recognize facial expression due to the complexity and subtlety of humanfacial expression. As far as the whole recognition system, the technology of facial expressionrecognition has got rapid progress. However, the performance of the facial expressionrecognition system is not sufficient to practical application and there are still many problemsneed to be further researched. One of the most important reasons is that facial expression willbe influenced by many factors, such as illumination, occlusion, pose and so on. In this paper,we put the emphases of the research on robust facial expression recognition undernon-uniform illumination and partial occlusion in order to improve the accuracy androbustness of expression recognition arithmetic in static facial expression image. The mainresearch contents and innovative work in this paper are shown as follows:First, in order to extract the texture information of facial expression and different scaleinformation of different expression behavior, we use Gabor filters to extract the facialexpression features. The Gabor multi-orientation fused features are combined with blockhistogram to extract facial expressional features in order to overcome the disadvantage oftraditional Gabor filter bank, whose high-dimensional Gabor features are redundant and theglobal features representation capacity is poor. In order to extract the multi-orientationinformation and reduce the dimension of the features, two fusion rules are proposed to fusethe original Gabor features of the same scale into a single feature. At the same time, torepresent the global features effectively, the fused image is divided into severalnon-overlapping rectangular units, and the histogram of each unit is computed and combinedas facial expression features. Experimental results show that the method is effective for bothdimension reduction and recognition performance.Second, a novel illumination-robust facial expression recognition method is proposed by using symmetric bilinear model to overcome the disadvantage of traditional2-dimensionalimage illumination preprocessing methods that they can degrade the quality of input imageand worsen recognition performance. We separate the illumination and expressioninformation which in the facial expression image under unknown illumination through thebilinear model, and build the illumination subspace and expression subspace respectively inorder to analyze and process the illumination information and expression informationindependently. The illumination factors are separated from the training database and theexpression factor is separated from testing image with arbitrary illumination, then the testingimage is transformed into a number of expression images exhibiting different illumination oftraining database. Experimental results show that the proposed method is better than thetraditional illumination preprocessing methods in recognition performance.Third, a novel facial expression recognition method under partial occlusion is proposedbased on local Gabor features radial grid encoding strategy in order to extract effective localfeatures to represent facial expression robustly. It has been found, in neurophysiological andpsychovisual studies, that two neighboring cells (both in retina and visual cortex) usuallyhave overlapping receptive fields. Therefore, in our implementation, a facial expressionimage is first divided into several local blocks which have50%overlap, and then each blockis represented by multi-scale and multi-orientation Gabor features, the resulting Gabor featureare encoded using radial grids, imitating the structure of human visual cortex. The proposedfeatures extraction method has the advantage of Gabor filters, which can represent the texturefeatures effectively and can overcome the disadvantage of Gabor filters, whose outputs arehighly correlated with redundant information at neighboring pixels. Better recognition ratesare achieved in JAFFE database with eyes occlusion and mouth occlusion. Experimentalresults show that the proposed local features coding method is effective to facial expressionrecognition under partial occlusion.Forth, an expression classification method based on support vector machine with localsummation kernel is proposed to overcome the disadvantage of conventional support vectormachine with global kernel, which can not process local features and is not robust toocclusion. The proposed recognition method based on local features, however, is robust toocclusion because partial occlusion affects only specific local features. In order to processlocal features in support vector machine effectively, local kernels are applied to process localfeatures and the the summation of local kernels is used as the integration method. Theeffectiveness and robustness of the proposed method are validated by comparison with globalkernel based support vector machine. The recognition rate is high under large occlusion, whereas the recognition rate of global kernel based support vector machine decreasesdrastically.Finally, the main content of this dissertation is summarized, and the further researchesare discussed.更多还原