The Research Of Facial Feature Extraction Method Based On Semi-supervised Learning

In the process of face recognition, feature extraction is the key step which determines the effect of recognition. A good feature extraction method probably leads to high recognition rate. The experts and scholars in the field of face recognition have proposed many feature extraction methods. Most of the classic feature extraction methods are only able to deal with the samples without supervised information or those with total supervised information. However, it’s not easy to get the supervised information, only a part of the samples have supervised information in most cases. Therefore, Semi-supervised learning methods are attracting more and more attention.This paper did some research on the face feature extraction method based on the core idea of semi-supervised learning. The creative work of this paper includes:(1) Semi-supervised Neighborhood Preserving Projection Based on Mahalanobis DistanceDimensionality reduction is a crucial step for pattern recognition tasks and finding a suitable low-dimensional subspace has an important effect on recognition performance. Recently, the computer vision and pattern recognition community has witnessed the rapid growth of the manifold learning. Among them, Semi-supervised neighborhood preserving projection (SSNPP) is one of the most promising algorithms. In SSNPP, neighborhood selection plays an important role. But there are some disadvantages in frequently-used neighborhood selection algorithm. For example, the value of k must be set in k-nearest-neighborhood (k-NN). In this paper, we develop algorithms by analyzing two key problems:1) The adaptive selection of the local neighborhood on the given set of high-dimensional data point.2) Apply the concept of Mahalanobis Distance to the neighborhood selection and promote the effect of manifold learning algorithm. The experimental results on YALE and ORL face image database show that the proposed method is effective.(2) Semi-supervised Dimensionality Reduction Based on Linear Local Tangent Space Alignment and Label PropagationIn the application of face recognition, Linear local tangent space alignment (LLTSA) isn’t able to take advantage of the sample label information. This paper improved LLTSA to propose a Semi-supervised dimensionality reduction based on linear local tangent space alignment and label propagation (SSLLTSA). Firstly, SSLLTSA gets the soft labels from the sample data with part of labels through the way of label propagation. Then, the soft label based scatter matrices are constructed to describe the inter-class separability and the intra-class compactness. SSLLTSA takes advantage of the information in the label while preserving the local structure of data. The experiments on YALE and ORL proved the outperformance of SSLLTSA based on traditional dimensionality reduction algorithms with maximum average recognition rate by 3.50% and 3.89% respectively.(3) Semi-supervised Two-dimensional Manifold Learning Based on Pair-wise ConstraintsConsidering the eigen-decomposition’s high complexity in one-dimensional Linear local tangent space alignment (LLTSA), this paper proposed a Semi-supervised two-dimensional manifold learning based on pair-wise constraints (2D-PCLTSA).2D-PCLTSA uses the samples of two-dimensional image matrices to extract image feature information, and adopts pair-wise constraints as supervised information.2D-PCLTSA takes advantage of the supervised information effectively while preserving the feature information in the sample set. Through the experiments on YALE and ORL,2D-PCLTSA outperforms based on traditional dimensionality reduction algorithms with maximum average recognition rate by 2.85% and 6.25%respectively. Especially, our algorithm could keep well classification performance with a few constraints.(4) Semi-supervised Dimensionality Reduction Based on Kernel Marginal Fisher Analysis and Sparsity PreservingConsidering the limitation that Marginal fisher analysis (MFA) can’t make use of the discriminant information in training samples, this paper proposed a Semi-supervised dimensionality reduction based on kernel marginal fisher analysis and sparsity preserving. Firstly, the proposed algorithm gets the sparse reconstruction of the samples. Secondly it uses the samples with labels to construct the inter-class ’penalty’ graph and intra-class ’similarity’graph. Then the algorithm gets the global information from all of the samples. At the end, we make it nonlinearized. The algorithm takes advantage of the information in both the unlabel samples and label samples. Experiments were conducted on YALE and ORL, our algorithm outperforms traditional dimensionality reduction algorithms with maximum average recognition rate by 2.48% and 4.88% respectively.