Research On Face Recognition Algorithms Based On Kernel Machine Learning

Over the past two decades, face recognition technologies have achieved great progress in practice. The facial images are always nonlinearly distributed due to complicated variations such as illumination, facial expression, pose and so on. Therefore, the performances of linear methods will be negatively affected. In order to extract nonlinear facial features, the kernel methods are widely used. Two typical kernel-based methods are kernel principal component analysis(KPCA) and kernel nonnegative matrix factorization(KNMF). However, these two algorithms are unsupervised learning approaches. They do not utilize the class label information of the training data, which would degrade the performance. In addition, the non-negativity of the implicit nonlinear mapping induced by the kernel function is not taken account into the current kernel based NMF face recognition algorithms. So, the kernel is nonnegative in-compatible. The objective of this dissertation is to remedy the aforementioned limitations.This dissertation includes five chapters. The first chapter introduces the related work. The conclusion and future work are given in Chapter five. The contributions are given in chapter 2 to chapter 4, which mainly involve the following three aspects.In the second chapter, a supervised KPCA(SK-PCA) is proposed for face recognition. By constructing a kernel matrix with the class label information, a nonlinear mapping is generated and well-defined on the training samples. Then this mapping is extended to the entire pattern space via interpolatory technology. Finally, we obtain an analytic expression of the nonlinear mapping. The function, created by the inner product of the nonlinear mapping, is theoretically proven to be a Mercer kernel function. Our supervised kernel is successfully applied to PCA method for face recognition. Two face databases, namely ORL and FERET databases, are selected for evaluations. The experimental results demonstrate that our SK-PCA has superior performance.As we know, the classical KNMF is an unsupervised method and then its accuracy will be degraded. Chapter 3 proposes a supervised KNMF(SKNMF) to overcome the above limitations of NMF. Based on the kernel theory and discriminant analysis method, this chapter establishes a new nonlinear NMF objective function incorporated with class label information. We obtain a new iterative algorithm for NMF. It theoretically shows that the objective function is non-increasing under the update rules. The proposed SKNMF method is successfully applied to face recognition. Five face databases are selected for evaluations. The experimental results demonstrate the superior performance of our method.Chapter four aims to tackle the non-negative incompatible kernel problem encountered in kernel based NMF. The existing KNMF methods cannot guarantee the non-negativity of the implicit nonlinear mappings. Therefore, the kernel functions are in-compatible under nonnegative constraints. In order to overcome this drawback, the fourth chapter proposes a novel method to construct nonnegative compatible kernel. Given a kernel matrix, we define a nonnegative nonlinear mapping on the training sample set using symmetric NMF technique. Then the mapping is extended to the whole sample space by means of non-negative interpolatory technology. It theoretically shows that the function, given by the inner product of the nonnegative nonlinear mapping, is a compatible Mercer kernel function. The proposed NC-Kernel is successfully applied to KPCA and KNMF. Three face databases, namely ORL, Pain Expression databases and CMU databases are selected for evaluations. Experimental results indicate our proposed NC-Kernel can enhance the classification power of algorithms.