Research On Several Algorithms Of Linear Projection Analysis And Its Applications

As the development of data acquisition technology in the last several years,the dimension of data feature space is becoming increasingly large.Many research areas have to face with the problem of handling those high-dimensional data which contain noisy and/or redundant feature inevitably,such as pattern recognition,computer vision,biology information and medical image processing etc.Therefore,many dimensionality reduction methods have been proposed for the purpose of solving-aforementioned problem.The main task of dimensionality reduction is to transform high-dimensional data into low-dimensional space while maintaining the structure information of original data as much as possible.Feature extraction technology is the most useful approach in the dimensionality reduction fields,and many researchers are in favor of the linear projection analysis algorithm.Numerous novel and improved linear projection analysis methods have been proposed such as least squares regression,metric learning and linear discriminant analysis etc.However,most of conventional methods suffer from several drawbacks or have limitations on some particular recognition tasks,such as non-orthogonal subspace,inaccurate nearest neighbors selection,Small Sample Size problem(SSS)and sensitive to outliers.In order to solve above problem,we propose some novel and improved algorithms and the main contributions of this paper can be summarized as follows:1.In order to restrain the deficient of non-orthogonal subspace in conventional LSR model.We propose a novel Orthogonal Least Squares Regression method(OLSR)which constructed under the orthogonal constraint and it can not only preserve more discriminant information in the subspace but also avoid the trivial solution.2.We propose Adaptive Neighborhood MinMax Projections(ANMMP)method for feature extraction.This algorithm iteratively assigns the optimal neighbors for each point adaptively so that can avoid the influence of noisy and redundant feature,then calculating the subspace more precisely.3.Motivated by the robust feature selection method,this paper proposes a new Robust Linear Discriminant Analysis(RLDA).It can not only overcome the SSS problem but also suppress the influence of outliers in the training data.