Research On Manifold Learning For Classification

In the classification task, an effective feature extraction method is an important prerequisiteto obtain accurate classification results. However, due to the huge quantity, high dimensionalityand nonlinearity of modern information, feature extraction have became a new challenge ofpattern recognition and machine learning. Traditional feature extraction has made great progressin dimensionality reduction, but there are still many problems to be solved in the field ofnon-linear and linear manifold areas. And then, the manifold learning attracts a surge ofreasearch interests as an effective dimension reduction method.This paper focuses on classic manifold learning and it consists of four parts:semi-supervised data, locality correlation, sub-pattern and noise data. The main innovative worksof this paper can be summarized as follows:(1)This paper proposes a novel approach to facial expression recognition based on labelpropagation techniques and the neighborhood preserving embedding. It considers both the globaland local structure of the training data. The approach first obtains a label matrix of the data byusing the classical label propagation algorithm, and then learns the manifold structure of the datain a low-dimension space. After that, it updates the label matrix in the feature space. As updatingthe label matrix takes consideration of the manifold characteristics of the data, the updated labelmatrix might be more accurate than the one obtained in the original space. Experiments onJAFFE dataset demonstrate the efectiveness of the approach.(2)This paper proposes a locality correlation discriminant with neighborhood preservingembedding for face recognition, which considers both the locality correlation and manifoldstructure of the training data. A new locality correlation preserving within-class scatter matrix isdefined, which not only contains the locality preserving information but also contains theneighbor correlation information, and defines a novel objective function to learn the manifoldstructure of the data in a low-dimensional space. Since the obtained manifold structure takesconsideration of the local neighbor correlation information and the discriminant information, itmight be more accurate for characterizing the feature of face images. Experiments conducted onYale and ORL database indicate the effectiveness of the proposed method.(3)Researches show that sub-pattern based face recognition approaches perform better thanwhole image based methods in local face information preservation. As manifold learningtechnologies preserve local manifold structure of the nonlinear sub-manifold while implementingdimension reduction, this paper puts forward a sub-pattern locality preserving projection(BspLPP). Unlike previous approaches that partition all training images of different classes intosub-images and use the same location images to form a sub-pattern, BspLPP first partitions thesame class images into different sub-images, uses the same location sub-images to form asub-pattern, and then applies LPP to learn the manifold structure of each sub-pattern.Experimental results show that BspLPP preserves the manifold and local information well andimproves the recognition performance.(4)The generalization performance of SVM applied to classification problems will be reduced if different class data are seriously overlapped. To cope with this problem, this paperpresents a new approach EB-SVM (Entropy Based Support Vector Machine) to pruning databased on the concept of information entropy for support vector machine. EB-SVM employs theinformation entropies of the training data to remove the patterns far from the boundaries anddelete the noise and overlapped instances closing to the boundaries, and then uses the pruneddataset to construct a SVM classifier. Experimental results show EB-SVM takes less time thanSVM and improves the classification accuracy.