The Research Of Subspace Analysis Methods Based On Semi-supervised Learning

Recently, semi-supervised learning and subspace analysis approaches are the most important research directions in the field of machine learning and pattern recognition subjects. Traditional supervised learning can only use a few labeled samples of training samples in the learning process, traditional unsupervised learning use lots of unlabeled samples of training samples in the learning process. Compared with traditional supervised learning and unsupervised learning methods, semi-supervised learning may simultaneously use labeled and unlabeled samples of training set, aiming at how to improve the generation ability of supervised learning and the efficient performance of unsupervised learning in the case of a small amount of labeled samples and a large number of unlabeled samples. Therefore, it is great significant that deeply research this important subject of semi-supervised learning.So far, lots of scholars have put forward many feature extraction measures. However, subspace analysis methods are widespread concern by lots of scholars owing to their appealing properties, such as strong description performance, small computational cost, easy to use and good separation and so on. At present, it has become a significant method for the research of face recognition. It is important to point out that face images are vulnerable to the interference of many external factors, many learns and researchers have been explored continually the issue that how to effectively obtain rich information from images.This paper mainly taking feature extraction of linear subspace analysis as the research object, focusing on semi-supervised learning as the research approach, carried out the following researches:(1) Adaptive Neighborhood Selection Semi-supervised Discriminant Analysis AlgorithmIn existing manifold learning methods are sensitive to parameters selection in constructing data local neighborhood and samples of category labels are inadequate in practical application of face recognition, this paper proposed an adaptive neighborhood selection semi-supervised discriminant analysis algorithm, named SSD-an. local neighborhood of samples by narrowing or expanding adaptively neighborhood coefficient to construct, effectively keeping the manifold local linear structure. At the same time, it also can take full advantage of oversight information contained in labeled samples, which can further excavate the hidden potential discriminant information in the unlabeled samples, and effectively maintain the samples of global and local information, so it has excellent capability of classification and recognition. Experiment results and analysis on FERET and ORL standard face database indicate that proposed SSD-an method completes effectively feature extraction and classification recognition, relative to the traditional feature extraction measures, enhancing the recognition rate, but also has more effectiveness. (2) Local Reconstruction and Dissimilarity Preserving Semi-supervised Dimensionality Reduction AlgorithmIn order to make the local feature and global feature between training samples applied effectively to subspace analysis approaches, a semi-supervised dimensionality reduction algorithm for feature extraction is proposed by combining local reconstruction with dissimilarity preserving, named LRDPSSDR. This algorithm can take full advantage of labeled samples and unlabeled samples of training set in learning process, which sets the edge weights of adjacency graph by minimizing the local reconstruction error and preserves local geometric structure of samples. In addition, dissimilarity between samples is represented by maximizing global scatter matrix so that global manifold structure between samples be maintained well. Experiment results and analysis in Yale and AR two standard face databases indicate that the proposed LRDPSSDR method has a better recognition performance and better robustness relative to the traditional feature extraction measures.(3) Semi-supervised Sparsity Preserving Two-dimensional Marginal Fisher Analysis Dimensionality Reduction AlgorithmFor the problem of labeled samples scarcity among samples sets, a smei-supervised sparsity preserving two-dimensional marginal fisher analysis algorithm is proposed, named SPP2DMFA. First, dimension reduction based on image matrix, spatial structure information of image pixels are effectively utilized. Then, it designs intra-class scatter matrix and inter-class scatter matrix in order to maintain the inter-class separability and the intra-class compactness according to labeled training samples. Finally, it constrains the sparse reconstruction among features by sparsity preserving, which not only preserves local geometric structure of samples, but also contains natural identification information of labeled training samples. Experiment results and analysis on AR, ORL and Yale standard face databases demonstrate that proposed SPP2DMFA method has excellent classification and recognition capability.