Research On Feature Selection Algorithm And Its Application In Image Recognition

Feature selection is of great significance in applications where data dimensionality is usually high,such as pattern recognition,text classification,and computer vision.It can reduce computational complexity and discover the inherent manifold structure of high-dimensional data.This paper studies the matrix regression model and the corresponding feature selection method,the main contents include:First,a robust graph-regularized sparse matrix regression(RGRSMR)model is given.Its loss function is based on the L_2,1norm,and the intra-class compactness graph based on manifold learning is used as the regularization term.RGRSMR combines intra-class compactness graph with sparse matrix regression to perform feature selection of two-dimensional matrix data.Secondly,a dynamic graph regularized label relaxed sparse matrix regression(DGRLR-SMR)model is established to realize the feature selection of two-dimensional matrix data.First,DGRLR-SMR takes matrix data as input data and introduces a non-negative label relaxation matrix,thereby relaxing the strict binary label matrix into a relaxation variable matrix to obtain discriminative information;Secondly,by dynamically learning the graph matrix,the model can also reveal the local geometric structure of the image sample.Therefore,the proposed method can obtain the discriminative information of the training data while maintaining the spatial position information of the original image data,and realize the feature selection of the image data through the sparse transformation matrix.Furthermore,a matrix regression model based on joint feature weighting and adaptive graph(FWGMR)is proposed.The model mainly learns a weight matrix and uses it to select some important features from image data.In addition,the graph embedding matrix is learned adaptively to reveal the precise local manifold structure of the training data.Therefore,the proposed method not only fully considers the spatial information and correlation of the elements in the image data,but also learns an accurate image embedding matrix,reduces the influence of noise,and improves the classification accuracy of the image.Finally,a margin-based discriminative embedding sparse matrix regression model(MDESMR)is proposed.For each matrix data,the difference between the two distances determined by the left/right regression matrix defines the interval between different classes.Maximizing the average interval of all training matrix data can obtain nonlinear discriminative embedding,enhance the difference between different classes,and further improve the efficiency of classification.Experiments on some public data sets show that the several algorithms proposed in the paper are feasible and effective,and have certain application value in image classification and recognition.