Application Of Feature Selection Algorithm In Image Classification And Recognition

Feature selection aims to find a set of simple features with good generalization ability by removing redundant,noisy and irrelevant features.The related algorithm has been widely used in the fields of bioinformatics,data mining and machine learning.Feature selection methods can be divided into supervised learning and unsupervised learning according to whether labels are used or not.This paper discusses the performance of different feature selection methods from unsupervised and supervised fields respectively.The main contents include:Firstly,an unsupervised feature selection（JURNFS）model combining uncorrelated regression and non-negative spectrum analysis is proposed.JURNFS selects uncorrelated and discriminant features and dynamically adaptively determines the similarity relationship between data,so as to obtain more accurate data structure and label information.Moreover,generalized uncorrelated constraints in the model can avoid trivial solutions,so this method has the advantages of uncorrelated regression and non-negative spectral clustering.Secondly,Nonnegative spectral clustering and adaptive graph-based matrix regression（AGNS-SMR）model are established for feature selection of unsupervised images to complete feature selection of two-dimensional unlabeled data.AGNS-SMR receives matrix data as input,thus saving the location information of the original matrix elements.The model can make the prediction label matrix as smooth as possible on the whole graph,and the graph weight matrix is learned through an adaptive process rather than a predefined fixed matrix.In this way,accurate local structure of sample data can be retained in the transformation space,so as to reveal the discrimination information of these pseudo-class labels.Furthermore,a supervised image feature selection（LNSRLP-MR）model based on low-rank non-negative sparse representation and locally reserved matrix regression is proposed for noise data.The model uses non-negative representation coefficients to adaptively learn the graph matrix,which can not only capture the global and accurate local structure information of the image data,but also has strong robustness to noise and occlusion of the image data.Specifically:LNSRLP-MR provides a non-negative sparse self-representation of each training image data by imposing non-negative and L₁-norm constraints on the representation coefficients,which makes the learned coefficients sparse and discriminant.In addition,the low-rank constraint of the error matrix is used to obtain the corresponding global structure information in the image data.Finally,an adaptive non-negative low-rank preserving sparse matrix regression model is proposed for supervised image feature selection（ANLRP-SMR）.The model first adopts a low-rank representation with non-negative constraints to capture the global structure and more discriminative information of the image data,and makes the error matrix of the self-representation of the training data tend to be sparse through the L₂-norm constraint.Secondly,a graph matrix learning model is established by combining the non-negative low-rank representation coefficients to reveal the local manifold structure of the image data adaptively.Therefore,the model can not only improve the robustness to noisy data,but also use the learned left and right regression matrices to obtain the corresponding row sparse transformation matrix to select features.Experimental results on some public data sets and scene data sets confirm that the feature selection algorithm proposed in this paper has certain value and superiority in the application of image classification and recognition.