Face Recognition Research Based On Subspace Analysis And Semi-supervised Learning

As a biological feature recognition technique, face recognition has been studying inpattern recognition and image processing field last decades and widely used in many fields,such as public security, identity authentication, financial information and public managementetc. Currently, there are many effective algorithms on face recognition, but no specificalgorithm can be used under all condition, because of problems such as dimensionalitydisaster, small sample size problem, variation of within-class scatter bigger than variation ofbetween-class scatter, complicated feature space and so on. Merging different featureextraction algorithms can effectively improve robustness and has wide application perspectiveand important theory value. Subspace analysis becomes the most important method on facerecognition with easy description and high effectiveness. To deeply study face recognitionmethods, the thesis is based on linear subspace analysis, and uses semi-supervised learning,support vector machine (SVM) kernel trick, discrete cosine transform (DCT), steerablepyramid transform (SPT). The main research work of thesis is as follows:(1) To solve high dimensional data dimensionality disaster and small sample sizeproblem of principal component analysis (PCA) and linear discriminant analysis (LDA) insubspace analysis, two semi-supervised DCT feature extraction algorithms are proposed,which are constrained clustering optimal discrete cosine transform (CCODCT) andsemi-supervised discriminant power analysis (SSDPA). In terms of frequency distributionproperty of DCT, premasking is used to reduce data dimensionality by selected low andmiddle frequency coefficients, so the computation cost is reduced. Based on constrainedinformation, CCODCT firstly perform semi-supervised constrained clustering on projectedDCT coefficients in order to improve recognition accuracy. Then compute discriminantcoefficient (DC) value of projected DCT coefficients. Finally, according to DC value, optimalmasking is selected and feature projection is performed. According to premasking, SSDPAcompute semi-supervised discriminant power (SSDP) value of projected DCT coefficients.Based on the SSDP value, feature projection is performed. The experiments show recognitionaccuracy of two algorithms outperform existing relative algorithms, and limitationdimensionality of small sample size problem of LDA is effectively solved.(2) To solve the problem which subspace analysis method is sensitive to illuminationvariation, a supervised DCT algorithm based on local region contrast enhancement is proposed. First, using contrast enhancement compensate for uneven illumination, convertoriginal image into contrast enhancement image in logarithm domain to enhance importantbut unobserved texture of image; then divide the contrast image into blocks, select importantlow frequency coefficient of each block in terms of DCT frequency property, merge them ascandidate feature vector to realized data reduction preliminary; compute block discriminantpower (BDP) value by labeled information; at last, in terms of BDP value, perform candidatefeature vector projection to extract effective feature. The experiment results show thealgorithm obtain higher recognition accuracy than other relative traditional methods with lowcomputation cost, effectively solve affect of uneven illumination and small sample sizeproblem, and improve the efficiency and accuracy of face recognition.(3) Most of subspace discriminant analysis methods have some limations, such ascommon covariance, uncertain dimensionality in feature space, low recognition rate and so on.A non-linear subspace feature extraction method, named Support vector direct discriminantanalysis (SVDDA) is proposed. The method incorporate support vector machine kenel trickinto subspace feature extraction, optimize the solution of eigenvalue problem; redesigndiscriminant criterion to solve small sample size problem. Thus, improve the face recognitionrate and robustness. With comparison with relative non-linear subspace kernel featureextraction methods, experiment results show SVDDA can obtain best performance, prove thealgorithm effectiveness and reliability.(4) Traditional wavelet transform approaches have high computation complexity, andcan’t decomposition in any direction and the limitation of sub-bands scale. A SteerablePyramid Semi-supervised Local Discriminant (SPSLD) algorithm is proposed. First, computestatistical information of each block of steerable pyramid sub-bands; then merge them into afeature vector in order to reduction data dimensionality; at last, in terms of small labeledinformation, perform semi-supervised local discriminant analysis to enhance recognitionaccuracy. With comparison with other global feature extraction approaches and othermulti-resolution wavelet transform approaches, the experiment results show SPSLD obtainbest performance and robustness for expression and illumination variation.