Feature Extraction Research Based On Discrimination Preserving Projection Approach

With the development of science and information technology, handling large-scale,high-dimensional data is a challenging task in modern. In recent years, many researchershave proposed various algorithms to solve the problem. Feature extraction is the one of themost effective and core technology.Feature extraction is an effective solution to the curse of dimensionality problem.Because of its excellent performance, it is widely used in machine learning，patternrecognition and bioinformatics. The goal of feature extraction is to construct a meaningfullow-dimensional representation of the original data. In general, feature extraction methods canbe divided into two categories: unsupervised ones and supervised ones.Linear subspace learning has achieved great success in feature extraction, and it aims tomap high dimensional data into low dimensional feature space which can reflect the importantinherent structure of the original data. In this paper, a novel approach termed DiscriminationPreserving Projection based on Sparse coding is proposed, which mainly focus on combininglocality supervised linear subspace learning with sparse coding. In our approach, wedecompose images into two parts including more discrimination part and less discriminationpart via dictionary learning and sparse coding firstly. Then, a locality supervised criterionwhich preserves the more discrimination part components while weaken the lessdiscrimination part components is presented. We conduct extensive experiments on publiclyavailable databases to verify the effectiveness and superiority of the proposed algorithm andcorroborate the above claims.