Research On Feature Subspace Learning Method Based On Representation Model

The two-dimensional image data,which is widely used in medical,transportation,security,and smart devices is the most intuitive form for humans to obtain information.Therefore,with the continuous increase of image dimensions,data dimensionality reduction has become a significant technology issue in computer vision.Among the many dimensionality reduction techniques,subspace learning has attracted wide attention and achieved great performance since it was proposed.Therefore,this paper focuses on feature subspace learning and proposes a series of feature subspace learning methods based on representation model.The specific research contents are as follows:(1)We propose a subspace structure constraints based feature subspace learning method via non-negative low rank representation,in which the low rank representation coefficient is used as the similarity weight to measure the distance of the pairs of samples in the projection space,and then the structural information of the sample is explored.Meantime,a label regression constraint is introduced to enhance the discriminant of the model and the applicability for classification problems.In addition,the joint learning of low rank representation coefficients and feature subspaces can benefit each other to achieve overall optimization.An ALM based solution algorithm is designed to guarantee the convergence of the objective function.Experimental results on different categories of datasets demonstrate the effectiveness of the proposed method.(2)We propose a low rank representation based robust feature subspace learning method,where a low rank reconstructed sample constraint term is designed to ensure the discriminant and robustness of the feature subspace learning model under various noise conditions,which can not only eliminate the noise interference in the data,but also use the label information to constrain the distance of the samples in the projection space.In addition,the proposed method also uses low rank representation coefficients to constrain the structural similarity of the feature subspace.In particular,different from the non-negative constraint on low-rank representation coefficients,we introduce a constant coefficient matrix to enforce positive and negative constraints on the elements in the low-rank representation matrix.Subsequently,an objective function solution algorithm is designed,and recognition tasks are performed on three face data sets under three kinds of noise conditions.Experimental results effectively verify the discriminant and robustness of the proposed method.(3)We propose a heterogeneous features based joint self-representation method for learning feature subspace and affinity matrix simultaneously,in which the two-dimensional spatial structure information of the multi-modal sample image composed of the original image and its multiple heterogeneous features is extracted,so that the spatial structure characteristics of the two-dimensional image will be effectively preserved and different modal samples will complement information.Then,the projected multi-modal data is used to construct a model based on shared ridge regression.In addition,motivated by the idea of graph learning,a non-linear structure constraint is designed to improve the ability for capturing data structures.Meanwhile,in order to effectively solve the objective function of the proposed method,a corresponding numerical algorithm is also designed.A solution algorithm for fast convergence of the objective function is designed,and spectral clustering experiments are performed on the face and object datasets.The experimental results indicate that the proposed method has outstanding performance.