基于非负矩阵分解的人脸识别算法研究

发布时间：2018-06-16 01:01

本文选题：人脸识别 + 非负矩阵分解　；参考：《深圳大学》2017年硕士论文

【摘要】：人脸识别技术在人们的生活当中扮演着重要的角色,它的关键在于特征提取.非负矩阵分解(NMF)是一种有效的人脸识别算法,它能够提取人脸模式的非负特征.然而,NMF是一种线性的无监督方法.由于人脸图像受光照、姿势和遮挡等因素的影响,它的数据呈现出复杂的非线性分布,此时,NMF的识别性能会降低.同时,NMF没有用到样本的类别信息,即它是无监督的方法,因此,它的分类效果将会受到影响.另外,NMF的收敛速度和特征稀疏度可以进一步被提高.最后,NMF不能进行增量学习,即当有新的训练样本或类别加入训练时,NMF需要对原来训练样本做重复的学习,这是非常耗时的.针对以上NMF算法的问题,本学位论文展开了深入的研究.本学位论文总共分为五章,第一章介绍了人脸识别背景和典型的算法,第二至四章讨论了本论文研究的主要工作,最后一章对研究工作进行了总结和展望.为了进一步提高NMF的收敛速度和识别性能,基于两种不同的误差测度,即欧氏距离和Kullback-Leibler散度(KL散度),在第二章,我们研究了两种快速非负矩阵分解算法(FNMF).在非负性约束下,FNMF在梯度下降法中选取了比NMF更大的迭代步长,这加快了NMF算法的收敛速度,并提取了更精确的特征.我们还可以证明,NMF是FNMF的一种特殊情况.为了进一步提高非负特征的判别力度,我们利用分块技巧到FNMF中,提出了两种分块的快速非负矩阵分解算法(BFNMF).BFNMF是一种线性的有监督算法,它具有一些良好的性质,例如它的特征是高度稀疏的且来自不同类别的特征是正交的.实验表明FNMF比NMF具有更快的收敛速度.同时,在ORL、FERET、pain expression(PE)和CMU PIE数据库上人脸识别实验还表明BFNMF取得了优越的识别性能,且FNMF比NMF取得了更好的识别效果.在第三章,我们研究出了一种分块的核非负矩阵分解算法(BKNMF),它是一种非线性的有监督方法.首先,通过利用每类的判别信息,我们构造了一个新的目标函数,它的目的是减小类内的距离.然后,基于核理论,通过求解目标函数的最小值,我们得到了BKNMF的迭代公式.最后,我利用分块技巧得到了BKNMF算法,并将它成功地应用于人脸识别.在理论上,我们证明了BKNMF算法的收敛性.BKNMF不但提升了NMF非负特征的判别力度,而且有效地提取人脸模式的非线性特征.在ORL、PE和Yale数据库上实验表明BKNMF取得了令人满意的识别效果.在第四章,我们研究出了一种基于分块的稀疏核非负矩阵分解(BSKNMF)的增量学习算法.在BSKNMF中,我们引入稀疏项和判别信息来构造了一个新的目标函数.基于该目标函数和核理论,我们利用分块技巧得到了BSKNMF.最后,基于BSKNMF,我们设计了增量学习算法,并将它成功地运用到人脸识别中.BSKNMF提取了比BKNMF更加稀疏和更有判别力度的特征,且来自不同类的特征也是正交的.在理论上,我们分析了BSKNMF的收敛性.当有新的训练样本或者类别加入训练时,我们的算法不需要对原来的训练样本做重复的学习,这极大地提高了时间效率.在ORL和Yale数据库上实验验证了我们算法的优越识别性能.
[Abstract]:Face recognition technology plays an important role in people's life, and its key lies in feature extraction. Non negative matrix decomposition (NMF) is an effective face recognition algorithm, which can extract non negative features of face pattern. However, NMF is a linear unsupervised method. Due to the illumination, posture and occlusion of the face image, the face image is unsupervised. At the same time, the recognition performance of NMF will be reduced. At the same time, NMF does not use the category information of the sample, that is, it is an unsupervised method, so its classification effect will be affected. In addition, the convergence speed and the characteristic sparsity of the NMF can be further improved. Finally, the NMF can not be carried out. Incremental learning, that is, when a new training sample or category is added to the training, NMF needs to repeat the original training sample, which is very time-consuming. In view of the problem of the above NMF algorithm, this dissertation has been studied in depth. This dissertation is divided into five chapters. Chapter 1 introduces the background of face recognition and typical algorithms, The two to four chapters discuss the main work of this paper. In order to further improve the convergence rate and recognition performance of NMF, based on two different error measures, namely, Euclidean distance and Kullback-Leibler divergence (KL divergence), in the second chapter, we study the two kinds of fast nonnegative matrix points. The solution algorithm (FNMF). Under the non negative constraints, FNMF selects the larger iteration step than NMF in the gradient descent method. This accelerates the convergence speed of the NMF algorithm and extracts more accurate features. We can also prove that NMF is a special case of FNMF. In order to further improve the discrimination of non negative features, we use the block technique to F. In NMF, two blocks of fast non negative matrix decomposition algorithm (BFNMF).BFNMF is a linear and supervised algorithm. It has some good properties, for example, its characteristics are highly sparse and the features from different categories are orthogonal. The experiment shows that FNMF has faster convergence rate than NMF. At the same time, ORL, FERET, pain expres The face recognition experiments on sion (PE) and CMU PIE database also show that BFNMF achieves superior recognition performance, and FNMF has better recognition results than NMF. In the third chapter, we have studied a block kernel non negative matrix decomposition algorithm (BKNMF), which is a nonlinear and supervised method. First, by using the discriminant information of each class, We construct a new objective function which aims to reduce the distance within the class. Then, based on the kernel theory, we get the iterative formula of BKNMF by solving the minimum value of the objective function. Finally, I use the block technique to get the BKNMF algorithm and successfully apply it to face recognition. In theory, we prove the BKNMF calculation. The convergence of the method.BKNMF not only improves the discrimination dynamics of the non negative features of NMF, but also effectively extracts the nonlinear characteristics of the face pattern. The experimental results on the ORL, PE and Yale database show that BKNMF has achieved a satisfactory recognition effect. In the fourth chapter, we have studied the increment of a block based sparse kernel nonnegative matrix decomposition (BSKNMF). In BSKNMF, we introduce sparsity and discriminant information to construct a new objective function. Based on the target function and kernel theory, we get BSKNMF. finally by block technique. Based on BSKNMF, we designed an incremental learning algorithm and successfully applied it to face recognition to extract more.BSKNMF than BKNMF. In theory, we analyze the convergence of BSKNMF. When a new training sample or class is trained, our algorithm does not need to repeat the original training sample, which greatly improves the time efficiency. In ORL and Yale data Base on experiments verify the superior recognition performance of our algorithm.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】