【作者】 李毅英；

【导师】 高西全； 高全学；

【作者基本信息】 西安电子科技大学 ， 通信与信息系统， 2011， 硕士

【摘要】 人脸识别作为一种典型的生物特征识别技术,以其自身的优势已经成为一个热点研究问题。在人脸识别中,特征提取是非常关键的一步,影响着后续分析和分类系统的性能。因此,如何有效地提取出反映数据本质结构的特征,方便后续分类,已成为目前需要解决的核心问题之一。在众多方法中,子空间分析是目前非常活跃的研究方向之一。本论文从子空间分析中的流形学习技术入手,深入研究了基于图论的图像空间几何结构描述,主要内容和贡献有:1.针对NPE不能较好地保持模式之间的差异信息,尤其是非线性数据的差异信息,提出了邻域结构保持嵌入(Neighborhood Structure Preserving Embedding, NSPE)。NSPE分别利用差异邻接图和相似邻接图描述数据空间的差异几何关系和相似几何关系,并给出了度量差异信息和相似信息的离散度矩阵;在此基础上,给出了一个意义明确、简洁的特征提取准则。2.针对NSPE需要将图像数据转换成向量,导致计算复杂和小样本等问题,提出了二维邻域结构保持嵌入算法(Two-dimensional Neighborhood Structure Preserving Embedding ,2DNSPE)。2DNSPE直接用图像矩阵来计算度量图象差异信息和和相似信息的离散度矩阵;在此基础上,给出了简洁、明确的特征提取准则。3.针对2DNPE和2DNSPE不能很好地保持图像像素的空间几何关系,导致冗余比较多和识别率不是足够的好等问题,分别提出了二维方向邻域保持嵌入(Directional Two-dimensional Neighborhood Preserving Embedding,Dir-2DNPE)和二维方向邻域结构保持嵌入(Directional Two-dimensional Neighborhood Structure Preserving Embedding, Dir-2DNSPE)算法。两种算法借助方向图对图象像素进行重组,有效地利用了像素的空间几何关系,然后分别利用Dir-2DNPE和Dir-2DNSPE提取图像特征,实验结果证实了所提方法的有效性。更多还原

【Abstract】 Face recognition is one of the representative biometric feature recognition techniques and has becoming an active research field in pattern recognition and computer vision. In face recognition, feature extraction is one of the very important key steps and involves the performance of the subsequent anslysis and classification system. So, how to effectively extract the intrinsic geometric structure embedded in data, which facilitates the subsequent analysis such as classification, has become one of the key problems to be solved. Many approaches have been proposed to slove it, and one of the most active approaches is subspace analysis methods (SAM). The dissertation mainly studies how to characterize the spatial geometric structures including similarity and diversity of data by combining graph theory and manifold learning belongs to SAM. The main contributions and work are as follows:1. Neighborhood Structure Preserving Embedding (NSPE) is proposed. NSPE defines two adjacency graphs, namely similarity graph and diversity graph, over the training data to model the spatial similarity and diversity structures of data, respectively. Similarity scatter and diversity scatter are calculated from the two graphs. Based on the two scatters, a concise feature extraction criterion is then raised by maximizing the ratio of the diversity scatter to similarity scatter.2. Two-dimensional Neighborhood Structure Preserving Embedding(2DNSPE) is proposed. 2DNSPE directly calculates the similarity scatter and diversity scatter from the image matrices. Based on the two scatters, a concise feature extraction criterion is raised by maximizing the ratio of the diversity scatter to similarity scatter. Different from NSPE, 2DNSPE avoids transforming the image matric into a vector, thus reducing the computational complexity and alleviating the small sample problem.3. Two novel methods, namely Directional Two-dimensional Neighborhood Preserving Embedding (Dir-2DNPE) and Directional Two-dimensional Neighborhood Structure Preserving Embedding (Dir-2DNSPE), are proposed to reduce dimension of images. Two methods rearrange pixels in images by using directional image and then performe Dir-2DNPE and Dir-2DNSPE to reduce dimension of images respectively. Experiment results show the efficiency of the two methods.更多还原