基于支持向量机的人脸识别技术研究

【作者】 付庆；

【导师】 赵志刚；

【作者基本信息】 青岛大学 ， 计算机应用技术， 2007， 硕士

【摘要】 人脸识别是生物特征识别的一个重要分支，也是计算机视觉与模式识别领域非常活跃的研究方向。相对于图像向量维数而言，人脸识别是一个高维、非线性小样本问题，而支持向量机解决该类问题，可以有效避免过学习现象，通过引入核函数，支持向量机将非线性不可分问题投射到高维空间后转化为线性可分问题。由于支持向量机最初是为解决两类分类问题提出的，如何将其应用于多类分类问题中仍然是目前研究的热点，本文提出一种改进的基于支持向量机的多类分类方法，有效地减少人脸识别过程的训练时间和测试时间并得到很高的识别率。本文主要做了以下工作：(1)分析比较基于支持向量机的多类分类方法，对构成分类边界的最优超平面进行分析，引入样本的特征空间距离，提出一种改进的基于支持向量机的多类分类器的构造方法，有效地减少分类超平面的数目，提高了训练效率，同时改进了基于“投票”机制和基于树的测试算法，并在UCI数据库上进行实验，与传统方法作比较；(2)研究了用于人脸识别的特征向量的选择和提取，使用基于小波和DCT的人脸特征提取方法，适当层次小波变换后的低频子图像刻画了人脸表情和姿势的不变特征，有较好的稳定性，由于图像能够将像块的能量集中于少数低频DCT系数上，因此提取经小波变换后人脸图像的DCT变换系数作为特征向量进行人脸识别；(3)在ORL人脸数据库上进行人脸识别实验，验证提出算法的有效性，人脸识别实验中从训练时间、最优超平面数目、测试时间和识别准确率等方面对几种基于支持向量机的多类分类方法进行了比较，证明本文的方法明显优于其他的方法，适合于人脸识别等实时性要求较高的应用。更多还原

【Abstract】 Face recognition is one of the important branches of biometrics and it is also one ofthe most active fields of computer vision and pattern recognition. Compared with thedimension of the face image vectors, face recognition is a high dimensional and nonlinearsmall-sample problem. Support vector machine (SVM) can solve it without over fittingphenomenon. By introducing the kernel function, the nonlinear separate samples areprojected into a high dimensional space (so called "the linear separate feature space") inwhich the new separate problem is solved. SVM is originally designed for binaryclassification. How to effectively extend it for multi-classification problems is still anattractive research issue. This paper proposes an improved multi-classification structurewhich helps save time of training and testing and has high recognition rate.In this paper our main work is as follows:First we analyze many multi-classification methods based on SVM and those hyperplanes contributing to classification boundaries. Then we propose an improvedmulti-classification structure considering distance between classes in the feature space.The new algorithm reduces the number of hyper planes, and thus improves the trainingefficiency. Meanwhile we have improved the testing methods based on max win and treemethodology. We carry out experiments on UCI database to compare the improvedmethod with traditional ones.Then we investigate feature selection and abstraction method for human facerecognition and put forward a method based on wavelet decomposition and discretecosine transform (DCT). The low frequency sub-images are obtained by utilizingtwo-dimensional wavelet transform are stable with features independent on faceexpression and pose. Only a small set of DCT coefficients is retained as the featurebecause most of image energy focuses on a few DCT coefficients.Finally experiments are carried out ORL human face database to verify validity ofour algorithm. We compare these multi-classification methods in the aspects of trainingtime, the number of hyper planes, testing time and recognition rate. The conclusion is ourmethod is fit for real-time problems such as human face recognition.更多还原