A Nearest Neighbor Subspace Preserving Feature Extraction Method And Its Application In High-Dimensional Images

With the rapid development of data acquisition technology,image data has become complicated and high-dimensional.Although high-dimensional images improve the performance of image recognition,they also increase the processing and storage costs of the recognition system.How to extract useful information from high-dimensional data is an urgent problem for researchers to solve.Among them,manifold learning is a widely used feature extraction method,which can capture and release the geometric information of data.Therefore,it has a good effect of dimension reduction.Locality plays an important role in manifold learning algorithms,but it is controversial to define locality in high dimensional space,especially when there is a lot of noise and redundant information in the data.Since it cannot define a locality in a true low-dimensional manifold,a practical method is to improve the locality confidence.The research content of this paper is as follows:(1)We analyze some typical feature extraction methods from linear and nonlinear perspectives,including their main ideas,derivation process,advantages and disadvantages.The linear feature extraction methods mainly analyze Principal Component Analysis and Linear Discriminant Analysis.The nonlinear feature extraction methods analyze Kernel Principal Component Analysis,Kernel Discriminant Analysis,Locality Preserving Projections,and Neighborhood Preserving Embedding.(2)To improve the locality confidence problem of manifold learning algorithms,we study and summarize some properties of the locality.As the same time,we propose a feature extraction method based on the Nearest Subspace Preserving.Firstly,every sample point and its nearest neighbors in the data are treated as a locality,and then a nearest subspace is stretched.Secondly,the Gram determinant is used to measure the volume of all the nearest subspaces.Finally,the volume is normalized and integrated into the model of the Locality Preserving Projections algorithm and Neighborhood Preserving Embedding algorithm.(3)Our experiments of clustering and classification run on real high-dimensional image sets,proving that the features extracted by our approach are more discriminative.