Research On Face Recognition Algorithms Based On Manifold Subspace Analysis

In the past few decades, along with advances in computer technology, as the hottest topicin the field of pattern recognition, face recognition has been widely discussed and studied.Feature selection and extraction is an important and difficult aspect in pattern recognition, andreducing the number of features is the important issue to design effective classifier in manycases. Face recognition is also true that we often transform the original features into fewernew features through mapping or transformation, and then face identification will be doneafter some characteristics with the best classification performance selected.So far, a large family of algorithms stemming from statistics or geometry theory has beendesigned to provide different solutions to the problem of face recognition. Overall, all of thesealgorithms can be divided into two categories, that is, linear and nonlinear. Among them,linear method has been widely used because its simple and low complexity time.The linear algorithms Principle Component Analysis (PCA) and Linear DiscriminantAnalysis (LDA) have been the two most popular because of their relative simplicity andeffectiveness. PCA searches for directions in the data that have largest variance andsubsequently project the data onto it. PCA is an unsupervised technique because of data type.However, LDA is a supervised algorithm and as such include label information of the data.But it should be noted that LDA algorithm is developed with the Gaussian distribution, aproperty that often does not exist in real-world problems. Without this property, separabilityof the different classes cannot be well characterized by interclass scatter. Another lineartechnique called Locality Preserving Projections (LPP) has been proposed for dimensionalityreduction that preserves local relationships within the data set and uncovers its essentialmanifold structure. There are lots of other methods, which pursue dimensionality reductionand image recognition in the literature, such as KPCA, KDA, ISOmap, Laplacian Eigenmapsand LLE.Nonlinear methods and kernel-based methods require more expensive computation forhigher recognition rate. PCA, LDA, LPP and MFA are simple in computational complexity.As the face is usually a low-dimensional manifold embedded in a high-dimensional space, wecan process primitive face data by means of kernel function, and to find the more reasonableneighbor points in the new feature space, rather than directly to calculate the Euclideandistance. In this way the constructed intrinsic graph and intraclass graph will be morereasonable and effective. In addition, although the MFA algorithm has largely improvedrecognition rate, it still has some typical problems such as singular value problems due to small sample size. In this paper we present a new discriminant function called differencecriterion to overcome the shortcoming of MFA algorithm and the classification will be moreeffective by selecting a appropriate parameter.The main work is as follows:Firstly, the background and status of face recognition are briefly described. Analyzes thedeficiencies existed in the subspace methods and gives the main contents and structuralarrangements of the paper.Secondly, some mainstream current manifold learning algorithms are introduced. Thebasic concept of manifold and manifold learning is given, and introduces several linear andnonlinear dimensionality reduction algorithms and feature extraction algorithms based kernelfunction.Next, we analyze the graph embedding framework and face recognition based on Fishercriterion.Then, we give a new face recognition algorithm called extended marginal fisher analysisbased on difference criterion. And we have done a lot of experiments to analyze itseffectiveness.Finally, we summarize the full text and look forward the future development direction ofthe research in this field.