The Research of Feature Extraction Methods and Their Applications
|School||Nanjing University of Technology and Engineering|
|Course||Pattern Recognition and Intelligent Systems|
|Keywords||Pattern Recognition Feature Extraction Fuzzy Set Principal Component Analysis Fisher Linear Discriminant Analysis Independent Component Analysis Small Sample Size Problem Face Recognition|
Face recognition is one of the hot topics in the field of pattern recognition,and it belongs to biometrics. In this field, feature extraction is one of thekey steps. In the passed decade years, many correlated algorithms have beenproposed to solve this problem. For example, linear discriminant analysis(LDA), principal component analysis (PCA) and independent component analysis(ICA) are developed to solve linear problem, and kernel methods based onsupport vector machine (SVM) are proposed to solve nonlinear problem. In thispaper, linear and nonlinear methods on feature extraction field are bothdeeply analyzed. Structure information is incorporated in the process offeature extraction method and more discriminant information is preserved inthe extracted feature space. Furthermore, effectiveness and performance areboth considered in our proposed method.Small sample size problem is a common problem on feature extraction fieldwhen Fisher linear discriminant analysis criterion is applied. Direct lineardiscriminant analysis (DLDA) is an effective feature extraction method, whichcan solve small sample size problem and extract optimal classificationfeatures from original samples. But conventional DLDA algorithm is oftencomputationally expensive and not scalable. In this paper, a new DLDA methodcalled DLDA/QR algorithm is proposed. With proposed method, computingcomplexity is reduced. Besides, effectiveness and stability of algorithm areimproved. At the same time, effective classification image space is definedin our paper based on image reconstruction. Then, effective features can beextracted from this space. Experimental results on ORL and XM2VTS demonstratethe effectiveness of the proposed method.Conventional linear discriminant analysis is based on binary classificationcriterion. That is to say, to be assigned sample, hard criterion is adoptedand the sample is classified to one class or not. While in real applications,samples are often affected by environment. For example, in the field of facerecognition, distribution of training samples is effected by expression,illumination et al.. In this paper, distribution of samples is representedby fuzzy membership degree. Therefore, sample can be classified into allclasses according to the current distribution. Furthermore, each sample’s contribution to classification can be represented by the corresponding fuzzymembership degree.Generally, difference of two classes is defined according to point-pointdistance in the process of feature extraction. In this paper, inspired bynearest neighbor line theory, a new feature extraction method called nearestneighbor line nonparametric discriminant analysis (NNL-NDA) is proposed.Furthermore, two preconditions should be satisfied with conventional Fishercriterion, which is very rigorous in real world applications. In our proposedmethod, nonparametric discriminant analysis is adopted, and influence ofthese factors is avoided.Kernel trick is developed based on support vector machine and statisticallearning theory. It is an effective strategy to solve nonlinear problem. Inthis paper, fuzzy linear discriminant analysis is generalized to fuzzy kernellinear discriminant analysis, and two-step strategy is used. Furthermore,conventional linear discriminant analysis should transform image matrix intothe corresponding vector. In our proposed method, two-dimensional lineardiscriminant analysis method is used to avoiding this transformation, andfuzzy membership function is incorporated. Experimental results on facedatabase demonstrate the effectiveness of this algorithm.Statistical independent component can be gained by independent componentanalysis (ICA) algorithm. Therefore, ICA plays important role in featureextraction field. Furthermore, ICA can be considered as the generalizationof PCA in high order. Similar to PCA algorithm, discriminant information isnot considered in ICA algorithm. In this paper, discriminant information isconsidered in ICA algorithm with Fisher discriminant criterion. LDA algorithmis incorporated into the ICA algorithm, and a new algorithm is proposed. Theproposed method not only gains the independent component, but also achievesdiscriminant information. Furthermore, symmetrical ICA algorithm is proposedto solve frontal face recognition problem.