基于代数特征的若干人脸识别方法研究

【作者】 胡静；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 计算机科学与应用， 2003， 硕士

【摘要】 人脸识别技术是计算机模式识别领域非常活跃的研究课题，在法律、商业等领域有着广泛的应用前景。由于人脸图像的特殊性，人脸识别问题也是模式识别领域的一个相当困难的问题，要使这一技术成为完全成熟的技术还有许多工作需要去做。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)、基于PCA的特征脸算法及特征值选择 本文从图像整体代数特征出发，首先介绍了“特征脸”算法的原理和实现过程。针对其在组成特征投影空间时特征向量选择问题的做了重点研究。本文从识别准确率和复杂度两个方面考虑，提出针对不同的样本集，选择不同的特征向量来组成特征空间，最后在ORL和Yale人脸数据库上进行实验。 (2)、基于Fisher线性判别的Subspace LDA人脸识别 由于PCA算法未能有效地利用训练样本类别信息，所以引入了基于Fisher线性判别的Subspace LDA人脸识别算法。本文介绍了Fisher线性判别准则原理和实现过程，然后引入Subspace LDA方法，用PCA将高维图像数据投影到低维的特征脸空间，再用LDA最大化判别系数。通过在ORL库和Yale库上的实验结果相比较，得出结论：由于LDA有效地利用了类别信息，消除了一些光照条件、表情、姿态的影响，降低了识别错误率，取得了优于PCA的效果。 (3)、基于逐对加权Fisher准则的改进LDA算法 应用Fisher准则可以得到“尽可能好的”分类结果，对应于准则函数取最优值。但是这个“尽可能好的’’分类结果对于它所对应的准则函数来说是最优的，但是相对于错误率或风险而言，它只是次优的。所以本文提出了基于逐对加权Fisher准则的改进LDA算法。该算法通过构建加权的类间散布矩阵，将距离较近的容易混淆的类别赋以较大的权值。并选择适当的加权函数，使得分类错误率逼近贝叶斯错误率。实验结果表明：基于逐对加权Fisher判别准则的人脸识别方法更加有利于人脸相似的特点，提高了人脸识别。更多还原

【Abstract】 The technology of face recognition is an active subject in the area of pattern recognition. There are broad applications in the fields of law, business etc. For the particularity of the face image, face recognition is also the very difficult problem. There is still much work to do. In this paper, some of face recognition algorithms are probed based on algebraic features of the images. And the corresponding solutions are given.The work including:(1) Eigenfaces based on PCA and the selection of the eigenvectorBased on algebraic features of the images, this paper first introduced the PCA-Based face recognition algorithm. We emphasized the selection of the eigenvector which used to create the eigenspace. Considering the recognition performance and the computation time, this paper proposed a method using to select the different number of the eigenvector in allusion to different training datasets. In the end, the ORL and Yale dataset are used for the experiments.(2) Subspace LDA algorithm based on Fisher linear discriminantThis paper introduced Subspace LDA algorithm based on Fisher linear discriminant for the PCA-Based face recognition algorithm can’t utilize the class information of train data. Firstly this paper introduced the principle of Fisher linear discriminant function, and then present the Subspace LDA algorithm which project the data that is in high dimension space to eigenspace that is in low space and then maximize discriminant coefficient. The result proved Subspace LDA algorithm eliminates some effection because of illumination or expression. This algorithm is better than PCA-Based face recognition algorithm.(3) Improved LDA algorithm based on pairwise weighted Fisher CriteriaThe application of Fisher function can have the best performance in classification relative to Fisher criterion. But this result is not the best result relative to Bayes error or risk. This paper introduces a face recognition method based on pairwise weighted Fisher Criteria. By constructing weighted between-class scatter matrix, the classes that are closer to one another are likely to have a greater confusion and should be given a greater weightage. Moreover, the classification error rate is related to the Bayes error by selecting appropriate weighting function. Experiments show that thenew method is useful in the classification because higher accuracy was achieved.更多还原