The Analyzing And Application Of Dimensionality Reduction Based On Manifold Learning And Subspace

The improving abilities of data collection and storage capabilities during the past decades have led to an information overload in most scientific domains. Traditional algorithms used in machine learning and pattern recognition applications are often susceptible to the well-known problem of the curse of dimensionality, which refers to the degradation in the performance of a given learning algorithm as the number of features increases. To deal with this issue, dimensionality reduction techniques are often applied as a data pre-processing step or as part of the data analysis to simplify the data model.Dimensionality reduction is the transformation of high-dimensional data into a meaningful representation of reduced dimensionality. It is important in many domains, since it applied in classification, visualization, and compression of high-dimensional data. When the suitable low-dimensional subspace is identified, which represent for the original high-dimensional data set through the dimensionality reduction transformation, tasks such as classification or clustering can often yield more accurate and readily interpretable results, while computational costs may also be significantly reduced.The dissertation presents a review and comparative study of techniques for dimensionality reduction at present. We also point out the disadvantage of these methods and then propose a novel dimensionality reduction method based on manifold learning and subspace, DRMS. It can preserve intra-class neighboring geometry structure and extract between-class relevant structures for classification effectively.The proposed method is applied to multimode biometrics recognition system with face and palm print images. The method is examined using the ORL and FERET face databases and PolyU palm print database. Experimental results show that DRMS consistently outperforms other linear dimensionality reduction methods when the training sample size of per class is small. This demonstrates the effectiveness and robustness of DRMS.