【作者】 晏哲；

【导师】 张莉；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 人脸是一个信息丰富的模式集合,是人类互相判别、认识、记忆的主要标志。人脸检测与识别在计算机视觉、模式识别、多媒体技术研究中占有重要的地位,因此人脸检测与识别技术是模式识别与机器视觉领域最有挑战性的研究课题之一。本文的工作涉及到了人脸的检测与识别,主要的工作内容如下：稀疏表示分类算法利用压缩感知的基本原理,通过求解由全部训练样本对测试样本最佳线性表出的稀疏向量来进行分类。在实验中发现,对于同方向分布的不同类的样本,稀疏表示分类算法在对样本单位化后无法正确分类。为了解决这个问题,本文将Mercer核引入到稀疏表示分类算法中,提出了核稀疏表示分类算法。由于高斯核函数可以作为样本间的相似性度量,这样就很好的解决了原算法出现的问题。在人工数据、UCI数据和人脸数据库上的仿真实验均验证了此改进算法的有效性。基于核稀疏表示的分类算法与稀疏表示分类算法中所应用的随机降维映射在实际的应用中是一种对维数约减非常有效的方法。但是对不同的随机映射降维矩阵,核稀疏表示分类算法会得到不同的识别结果,因而不能够保证分类算法的稳定性。如果想提高算法的性能,对不稳定的分类器而言,集成是一种很好的选择。因此我们提出了基于核稀疏表示分类算法的多分类器集成方法,可采用的多种决策级融合方法包括：最大(Max)、最小(Min)、求和(Surn)、均值(Mean)和多数投票(Majority Vote)。实验验证了该集成方法的有效性,而且实验表明了求和与均值这两种策略是较好的集成规则。标准支持向量机(SVM)已经用于人脸检测,但是由于支持向量个数较多,导致检测速度不高。1-norm SVM采用1-norm正则项替代了标准SVM中的2-norm正则项,而1-norm正则项能够诱导稀疏性。已经证明了1-norm SVM的解比标准SVM的解更具有稀疏性,因此我们把1-norm SVM应用到人脸检测中,期望能提高检测速度。在构建1-norm SVM人脸检测系统的时候,采用的是经典的人脸检测系统,并在最后加入了去除重叠标识人脸的步骤。最后通过实验验证了1-norm SVM的确能够提高检测速度,而且检测精度和SVM可以相媲美。更多还原

【Abstract】 Human face embodies extremely rich information and is the key symbol for distinguishing, recognizing and memorizing individuals. It plays an important role in computer vision, pattern recognition and multimedia technology. Therefore, automatic recognition of human face is one of the most challenging subjects in pattern recognition and computer vision. Our work in this thesis relates with two areas of automatic recognition of human face:face detection and recognition. The main work is as follows.Based on the theory of Compressed Sensing (CS), the goal of Sparse Representation Classifier (SRC) is to find a sparse vector which is a linearly optimal representation for a test sample by using all training samples, and use it to classify this test sample. If the samples are distributed in the same direction, SRC can not classify them exactly. To solve this problem, we propose a Kernel Sparse Representation Classifier (KSRC), which introduces the Mercer kernels to SRC. As the similarity measure between samples, RBF kernel function is a good solution to the problem. The experiments on artificial data, UCI database and Extended Yale B database have verified the effectiveness of KSRC.Based on Kernel Sparse Representation Classifier, we present an ensemble method for KSRCs. The random projection used in KSRC and SRC is an effective way for dimensionality reduction. But for different random matrix, KSRC will get different results. So we use the ensemble of KSRCs to ensure the stability of the algorithm. There are many rules to combine the multiple classifiers, such as Max, Min, Sum, Mean, and Majority Vote rules. The experiments have verified the effectiveness of combining classifiers, and also show that Sum and Mean are better ways.Support Vector Machines (SVMs) have been applied to face detection. But the test speed of SVMs is not satisfied. In order to improve the test speed, we apply 1-norm SVMs to face detection.1-norm SVMs adopt the 1-norm regularization which can induce sparsity. It has been shown that the solution of 1-norm SVM is sparser than standard SVM, so use 1-norm SVM to face detection can improve the detection speed. We have verified the effectiveness of reducing the detection time by experiments.更多还原