【作者】 隋广洲；

【导师】 王金城；

【作者基本信息】 大连理工大学 ， 模式识别与智能系统， 2007， 硕士

【摘要】 人脸识别是近年来模式识别和图像处理领域的研究热点之一，对该问题的研究有助于模式识别和信息安全的发展。而特征抽取是模式识别研究的最基本问题之一，特征抽取方法对于图像识别起着关键的作用。在人脸识别领域，基于代数特征的人脸识别方法因其计算简单、有效等特性引起了人们的广泛注意，现已成为人脸图像特征提取和识别的主流方法之一。本文结合几种基于代数特征的人脸识别算法，对其中的部分问题分别进行了探讨，并给出了相应的解决方案。本文工作包括：(1)对基于奇异值分解(SVD)、主成分分解(PCA)、Fisher线性鉴别分析(FLD)以及二维主成分分解(2DPCA)理论做了详细的分析和介绍。这几种方法是人脸识别领域广泛运用的基于代数特征的特征提取和识别的方法。经过长期不断的试验证明尽管每种方法都有其独特的优越性，但同时每种方法也都具有各自的弊端，仍然达不到明显的理想的识别效果。本文利用试验给出了各种方法在ORL人脸库中的识别结果。分析比较了各种方法的特征提取的优劣性。(2)在分别介绍了SVD和PCA的人脸特征提取方法之后，提出了SVD和PCA相结合的人脸识别算法。理论上，当两种数据或分类器具有一定的独立性或互补性时，数据融合或分类器融合才能改善识别率。SVD和PCA之间有着明显的互补之处。PCA在图像表示上是最佳的(在均方差意义上)，但敏感于位移、旋转等几何变换。而SVD则具有位移、旋转不变性。因此，将这两种方法相结合就有可能提高分类性能(好于单独的SVD方法和单独的PCA方法)。在ORL数据库上的实验表明，SVD和PCA相融合的识别方法的确提高了人脸识别率。(3)2DPCA方法是在PCA方法的基础上提出的，具有比PCA方法有更多的优点，识别率更高。FLD方法更清楚的体现了各类别之间及类内部各元素间的关系，使提取的信息更具有独立性，根据FLD这个特点本文在利用2DPCA进行特征提取的过程中使用了FLD中的类间散布矩阵提出基于类间散布矩阵的2DPCA人脸识别算法，并且根据加法融合理论提出了FLD+2DPCA识别算法，由两种融合算法期望达到更良好的识别性能，试验结果证明所提出的方法达到了预期的融合效果。更多还原

【Abstract】 Face recognition is hot in the area of pattern recognition and image process recently, and the research for it will be benefit for the progress of pattern recognition and information security. Feature extraction is one of the foundational problems in the research of pattern recognition, and it is the key to the problem of image identification. In the field of face recognition, the method based on the algebraic features of the images has been received extensive attention owing to its easily computation and effectiveness. Now the face recognition algorithms on the basis of algebraic features of the images has become the mainstream technology for feature extraction and face recognition. In this paper, combined with several methods of face recognition algorithms based on algebraic features of the images, some problems in them has been probed, and the corresponding solutions are given.This paper includes the following parts:(1) There are brief introduction and clearly analyses of the theories of SVD、PCA、FLD and 2DPCA, which are used widely in face recognition that based on algebraic features. Through the experiments again and again, we find although every method has it’s own advantages but also has some disadvantages, so it can’t get the perfect result. In this paper, the testing results of each method are given after the experiments of recognition in ORL face image base, and the advantages and disadvantages of each method are compared.(2) After introducing the feature extraction methods of SVD and PCA respectively, a face recognition method based on the integration of SVD and PCA is put forward. Theoretically, fusion of different data or classifiers can achieve better performance when they are independent or complementary. SVD and PCA are clearly complementary to one another. PCA is the best one in image expression, but one of drawbacks of PCA-based method is that PCA is sensitive to translation, rotation and other geometric transforms. Contrary to PCA, SVD has the merit of invariance to translation, rotation and other geometric transforms. By combining these two methods, it is expected that better recognition performance can be obtained. Experiment results from ORL face database demonstrate that the proposed method can indeed improve face recognition rate.(3) 2DPCA is proposed based on PCA and it have more advantages than PCA for it’s higher capacity of face recognizing. FLD makes the relation of the classes and the relation of the elements in one class clearly and it also make the elements extracted more unattached. According to this character of FLD, this paper has fused the method of FLD into the 2DPCA in the process of extracting the features, and also shows another method called LDA+2DPCA, hoping to achieve a far more effective capacity of recognition. The result of the test shows that the methods being introduced have achieved the expected fusion effect.更多还原