基于PCA和SVM的人脸识别关键技术研究与实现

发布时间：2019-03-05 11:29

【摘要】：人脸识别作为当今科研领域最热门的课题之一,它的应用已经深入到了金融系统、信息安全、公共安全等各个领域。横向对比来看,人脸识别在众多模式识别方式中优势明显,具有自然性、非侵扰性、采集成本低以及人机交互性强的天然优势,应用场景更加广阔。纵向分析,各大上市公司不断推出大量商业化人脸识别产品,人脸识别将在未来呈现爆发式的增长。因而研究人脸识别具有重大的现实意义。本文重点研究了人脸预处理、人脸特征提取和人脸识别这三大内容。尤其是对主成分分析(Principal Component Analysis,PCA)和支持向量机(Support Vector Machines,SVM)算法进行了深入的研究,并针对其并未利用类别信息的不足,引入了线性判别分析(Linear Discriminant Analysis,LDA),提出了基于PCA和SVM的人脸识别改进框架,并基于改进后的框架,实现了一个实时人脸识别系统。本文的主要工作有以下五个部分。(1)研究了多种常用的人脸图像预处理技术,包括彩色图像的灰度化、图像的灰度变化、直方图均衡化及几何归一化等方法。人脸图像预处理技术降低了外界因素,如光照、姿态、拍摄角度等问题对图像的影响,实现了图像的标准化处理,为之后的人脸识别工作的开展奠定了良好的基础。最后,利用Matlab做了相关的算法检验工作。(2)深入研究了基于PCA的人脸识别的理论基础和具体的实施过程。由于在PCA算法中直接计算协方差矩阵的特征值和特征向量计算量太大,所以引入了SVD定理,实现了主元子空间的间接运算。并利用Matlab进行了大量的实验,对PCA算法的优劣进行了分析,得出PCA算法具有良好降维效果,但并没有分类的效果。(3)深入研究了线性支持向量机的理论基础,论述了SVM的分类器在小样本以及学习等方面的优点。然后运用核函数,将线性支持向量机扩展成非线性支持向量机,使得支持向量机能够满足人脸识别的分类要求。讨论SVM多分类方法的实现。(4)针对PCA算法中并未利用分类信息的问题,引入线性判别分析,并提出基于PCA和SVM的人脸识别改进框架,最终形成了PCA+LDA+SVM的人脸识别框架,并在Matlab平台上经过大量的实验,探讨了维数比率、训练样本数目对算法的影响,验证了它良好分类效果。(5)基于改进的人脸识别框架,用Visual Studio 2010、Open CV库以及Qt框架实现了一个实时人脸识别系统。通过动态人脸识别的测试,识别率为97.3%,识别效果优良,每次的识别时间平均为356毫秒,满足实时性的要求。
[Abstract]:Face recognition is one of the most popular topics in the field of scientific research. It has been applied in many fields, such as financial system, information security, public security and so on. In horizontal comparison, face recognition has obvious advantages in many pattern recognition methods, such as naturalness, non-intrusiveness, low collection cost and strong man-machine interaction, so the application scene is wider. Longitudinal analysis, the major listed companies continue to launch a large number of commercial face recognition products, face recognition will show explosive growth in the future. Therefore, the study of face recognition has great practical significance. This paper focuses on face preprocessing, face feature extraction and face recognition. Especially, the principal component analysis (Principal Component Analysis,PCA) and support vector machine (Support Vector Machines,SVM) algorithms are studied in depth, and the linear discriminant analysis (Linear Discriminant Analysis,LDA) is introduced to solve the problem that the class information is not utilized. An improved framework of face recognition based on PCA and SVM is proposed, and a real-time face recognition system is implemented based on the improved framework. The main work of this paper includes the following five parts: (1) A variety of pre-processing techniques of face image are studied, including grayscale of color image, gray level change of image, histogram equalization and geometric normalization and so on. Face image pre-processing technology reduces the influence of external factors, such as illumination, pose, shooting angle and so on, and realizes the standardized processing of face image, which lays a good foundation for the development of face recognition work. Finally, the related algorithms are tested by Matlab. (2) the theoretical basis and implementation process of face recognition based on PCA are deeply studied. Because the computation of eigenvalues and Eigenvectors of covariance matrix is too large in PCA algorithm, the SVD theorem is introduced to realize the indirect operation of principal subspace. A large number of experiments are carried out with Matlab, and the advantages and disadvantages of PCA algorithm are analyzed. It is concluded that PCA algorithm has good dimensionality reduction effect, but it has no effect of classification. (3) the theoretical basis of linear support vector machine is deeply studied. The advantages of SVM classifier in small samples and learning are discussed. Then the kernel function is used to extend the linear support vector machine to the nonlinear support vector machine, so that the support vector machine can meet the classification requirements of face recognition. The implementation of SVM multi-classification method is discussed. (4) aiming at the problem that the classification information is not used in PCA algorithm, linear discriminant analysis is introduced, and an improved framework of face recognition based on PCA and SVM is proposed. Finally, a face recognition framework of PCA LDA SVM is formed. Through a large number of experiments on Matlab platform, the influence of dimension ratio and number of training samples on the algorithm is discussed, and its good classification effect is verified. (5) based on the improved face recognition framework, Visual Studio 2010, A real-time face recognition system is implemented based on Open CV library and Qt framework. Through the test of dynamic face recognition, the recognition rate is 97.3%, and the recognition effect is good. The average recognition time is 356 milliseconds each time, which meets the real-time requirement.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】