【作者】 梁万杰；

【导师】 赵建民；

【作者基本信息】 浙江师范大学 ， 计算机软件与理论， 2009， 硕士

【摘要】 人脸识别方法和关键技术是当前模式识别和计算机视觉领域的一个研究热点。人脸识别的步骤主要包括人脸检测、特征提取和特征分类。本文主要对人脸特征提取和特征分类进行了深入的探讨和研究,提出了基于Gabor小波变换的有监督局部保持投影算法(Based on Gabor Feature Supervised locality preservingprojection,GSLPP)并应用于人脸特征提取,及最近邻特征线分类器与支持向量机分类器二级融合的分类方法应用于特征分类,本文的主要工作如下:(1)在人脸识别的特征提取阶段,本文采用了基于Gabor小波的一种有监督的LPP算法,该方法先用Gabor小波对人脸图像进行滤波处理,然后直接对每个LPP基向量进行线性判别分析,选择最具判别力的基向量来构造子空间。Gabor滤波器对于图像的亮度和对比度变化以及人脸姿态变化有较强的鲁棒性,并且它表达的是对人脸识别最为有用的局部特征。基于以上优点本文先采用Gabor小波对人脸图像进行滤波处理。提取人脸Gabor特征通常要对一幅图像进行四十次滤波处理,这会使图像的维数增加,从而影响识别的效率,为了提高识别的速度,必须采取有效的降维方法来降低数据的维数。本文采用有监督的局部保持投影算法提取经Gabor小波滤波后的人脸图像的主要特征。局部保持投影是一种新的子空间分析方法,它是非线性方法Laplacian Eigenmap的线性近似,既解决了PCA等传统线性方法难以保持原始数据非线性结构的缺点,又解决了非线性方法难以获得新样本点低维投影的缺点。但LPP算法是一种无监督的学习方法,当人脸图像的光照、姿态、表情发生变化时,LPP的识别率会迅速下降。本文提出的一种有监督的LPP算法,因其直接对每个LPP基向量进行线性判别分析,选择最具判别力的基向量来构造子空间,有效的提取了人脸图像的分类特征,实验证实其在特征提取方面具有较好的效果。(2)在人脸识别的人脸特征分类阶段,本文提出最近邻特征线分类器与支持向量机分类器以串联的方式构建一个二级融合的分类器。如果将多种不同的分类器以某种方式进行组合,就有可能在总体上取得比单一分类器更好的分类效果。基于此观点,本文提出了一种将最近邻特征线分类器与支持向量机分类器进行二级融合的分类器设计方法。利用该融合分类器分类时,先采用最近邻特征线分类器进行第一级分类,若结果大于设定的阈值,则拒识,否则,转入后级分类器,用支持向量机分类器进行精确分类。该多分类器融合方法不仅充分利用了支持向量机分类器识别率高和最近邻特征线分类器速度快的优点,而且还利用最近邻特征线分类器的结果指导支持向量机分类器的训练和测试,从总体上提高了分类的精度。应用上述方法,结合FERET人脸图像库和JDL人脸数据库进行实验,结果表明本论文提出的人脸识别方法与其它(PCA,LDA等)识别方法相比具有较好的人脸识别效果。更多还原

【Abstract】 Face recognition is a focus of the research in the field of Pattern Recognition and Computer Vision which mainly includes face detection, facial feature extraction and facial feature classification. This paper discuss mainly on facial feature extraction and facial feature classification. This paper put forward a facial feature extraction algorithm which base on Gabor wavelet feature using supervised locality preserving projection, and a multi-classifiers fusion method which combine nearest feature line classifier and support vector machine classifier. The main contents of the paper can be noted as following:(1) At the facial feature extraction stage, this paper proposes one facial feature extraction algorithm which base on Gabor wavelet feature using supervised locality preserving projection and is abbreviated GSLPP. Due to the good characteristic of human face image’s Gabor wavelet feature, face recognition technology base on Gabor wavelet feature is a very popular method. However, human face image’s Gabor filter processing will increase the dimension of data. In order to reduce the dimension of data, this paper use supervised locality preserving projection algorithm to reduce the dimension of data. Locality preserving projection algorithm is a new subspace analysis method, which can preserve the most important part for face recognition. The method was successfully applied to face recognition in controlled environment. But the locality preserving projection algorithm is a non-supervised study method. When the human face image’s illumination, posture and expression changed, the recognition rate will decrease dramatically. For improving the recognition rate, this paper put forward a supervised locality preserving projection algorithm.(2) Based on the view that combining some kinds of classifier by some way maybe obtain higher recognition rate than a single one, this paper put forward a multi-classifiers fusion method which combine nearest feature line classifier and support vector machine classifier. The step of the multi-classifiers fusion method as following: firstly, using the nearest feature line classifier to pre-stage classify, if the result bigger than the defined threshold, then rejecting recognition, else pass to SVM classifier to classify. The multi-classifiers fusion method not only makes full use of the advantage which is high recognition rate of SVM classifier and high recognition speed of the nearest feature line classifier,but also utilizes the classification result of the nearest feature line classifier to guide the training and classification of SVM classifier.At last, the paper test this human face recognition method by some experiments on FERET Face Database and JDL-A Face Database. Those experiments illustrate that the method proposed by this paper is more effective than other methods, such as PCA、LDA and so on.更多还原