Dimension Reduction Algorithm And Its Application In Image Recognition

As a data preprocessing process,dimensionality reduction plays an important role in data mining,pattern recognition and machine learning.Using dimension reduction algorithm technology can reduce the complexity of the problem,so as to improve the prediction accuracy,robustness and interpretability in machine learning calculation.This paper focuses on the dimensionality reduction algorithm supported by a large number of theoretical research and repeated experiments.Firstly,a robust unsupervised feature selection（FSRGR）model based on feature self expression and graph regularization is proposed.The model uses feature self expression which linearly represents each feature with other features.Secondly,the row sparse weight matrix can be obtained by using the graph regularization term based on L_2,1-norm to reduce the influence of noise on the data.Finally,the low rank constraint is imposed on the weight matrix,which not only retain the local data structure,but also reveal the global structure.Secondly,a low-rank approximation-based two-directional linear discriminant analysis for image recognition method（LRAM+2DLDA）is proposed.Specifically,a set of low rank matrices are used to approximate the original data matrix,and the characteristic matrix of the training image is obtained through the low rank matrix,which is used to define the inter class scattering matrix and the intra class scattering matrix.This method minimizes the reconstruction error of the transformed characteristic matrix,and achieves the goal of minimizing the intra class dispersion and maximizing the inter class dispersion.Furthermore,a sparse low rank approximation matrix and local preserving model（SLRAM-LP）are proposed for unsupervised image feature selection.The model takes the image matrix data as the input,avoids the high computational complexity of high-dimensional vectorization data,and uses a set of low rank matrices to approximate the training image data matrix.Then uses the Kronecker product of two low rank transformation matrices and L_2,1-norm sparse regularization to avoid cumbersome solution and realize the feature selection of the image.Finally,The model also preserves the local structure of the original sample matrix into the transformation space by minimizing the weighted distance between the simplified representations of the original sample matrix.Finally,a flexible sparse robust low rank approximation matrix model（FSRGLRAM）is proposed,which combines feature selection with subspace learning and eliminates redundant features.This method is different from the previous low rank approximation model,and introduces two recovery matrices.L₁-norm is applied to the reconstruction error and L_2,1-norm is applied to the Kronecker product of the left and right transformation matrix to reduce the influence of noise on the reconstruction error,and feature selection is carried out while learning the optimal transformation matrix and recovery matrix.The experimental results on some public data sets show that the above four algorithms are effective and feasible in the application of image dimensionality reduction.