【作者】 刘翠响；

【导师】 孙以材；

【作者基本信息】 河北工业大学 ， 微电子学与固体电子学， 2008， 博士

【摘要】 人脸识别近年来在模式识别领域受到高度关注。作为一种主要的生物识别技术,在金融安全、电子商务与数字娱乐等领域具有广泛的应用前景。经过近半个世纪的发展,在人脸识别研究领域已经取得许多成果,基本实现了特定环境下的准确识别。尽管如此,人脸识别技术要达到完全实用水平,还面临着诸多挑战,因人脸数据维数过高带来的大规模数据存储和计算问题就是其中之一。本文针对人脸识别中的高维数据降维及识别问题展开深入研究,主要研究工作与创新点如下:1)研究了基于典型线性与非线性降维算法的人脸图像高维数据的降维问题。深入分析了不同算法的降维性能;基于剩余残差的维数评价模型,对人脸图像的本征维数进行了讨论;比较了Isomap算法的不同邻域参数k对降维结果的影响,并在ORL、Yale、Feret人脸库上进行对比试验。试验结果表明,非线性降维算法Isomap的降维效果优于典型的线性降维方法PCA。2)研究了基于非线性降维的人脸识别方法。为了解决新样本在训练空间的映射及降维问题,引入了增量式Isomap算法,提出了利用距离保持的增量式(IADP-Isomap)的人脸识别方法,首先对人脸图像应用Isomap,然后采用IADP-Isomap获得新样本的低维特征表示,采用最近邻分类器进行分类。实验结果展示了该方法的可行性。3)探讨了非负矩阵分解(NMF)算法的原理及在人脸识别中的应用。基于NMF的人脸识别方法在人脸图像的光照、姿态与表情改变时性能会大幅下降,针对该问题,本文提出NMF+SDA方法对之改进,首先对人脸图像应用NMF,然后融合线性鉴别分析(LDA)和奇异值分解(SVD)的思想,进行人脸高维数据降维和特征提取。在人脸库上的实验表明,本方法具有较高的识别率。4)基于非线性降维算法,引入模糊数学中矢量隶属函数和隶属度,提出了一种基于模糊隶属函数的三控制要素的多项式模糊拟合算法。利用归一化贴近度可以评价拟合曲线的线性度。将此算法用于人脸识别,实验结果表明具有较好的识别率。更多还原

【Abstract】 In recent years, the face recognition has attracted considerable attention within the community of pattern recognition. As one of the most successful branches of biometrics, it has great potential applications in finance security, electronic commerce, and digital entertainment, etc. Over the past half of century, the face recognition has developed rapidly. Now under the controlled conditions, face recognition systems have achieved good results. However, a great number of challenges are still leaved to resolve before one can implement a robust and practical face recognition application. Among these challenges, the large-scale data storage and computation arising from excessively high face data is one of the most difficult.Our work is focusing on the dimension reduction of face data and recognition problem. The work and the innovation in this dissertation can be summarized as following.(1) The dimensionality reduction problem of face data based on typical linear and nonlinear dimensionality reduction algorithms is investigated. Meanwhile, the performance of these algorithms is intensively analyzed. Based on the residual variance evaluation model, this dissertation discusses the intrinsic dimension of facial images. Following that, the influences of neighborhood parameters k on dimensionality reduction are taken into account. Experiments on the ORL, Yale, and Feret face database show the performance of nonlinear dimensionality reduction algorithms is better than that of linear ones.(2) The face recognition method based on the nonlinear dimensionality reduction algorithm is studied. Firstly, this dissertation introduces the incremental Isomap algorithm to resolve the novel samples’mapping and dimensionality reduction problem in the training space. Following that, a face recognition method based on IADP-Isomap is proposed. The experimental results show that the recognition method is feasible.(3) The non-negative matrix factorization (NMF) algorithm and its application in face recognition is discussed. On condition of the variation of illumination, poses, and expression, the performance of NMF-based recognition method would dramatic decreases. Focusing on this problem, this dissertation proposes a so-called NMF+SDA algorithm. It can effectively implement dimensionality reduction and feature extraction of the face dada. Experiments on face database exhibit that NMF+SDA owns better recognition rates than traditional NMF.(4) Based on the nonlinear dimensionality reduction algorithm, the concepts of vector membership function and membership degree in fuzzy mathematics are introduced. It is presented that the fuzzy matching for a nonlinear function between input and output can be realized by using three rulers (two point rulers and one slope ruler). The affinity between two memberships can be used for assessment to the linearity of the matched curve. Consequently, the algorithm of polynomial fuzzy matching based on three rulers is proposed and applied in face recognition. Experimental results demonstrate the recognition algorithm is feasible and has good recognition capability.更多还原