Research On The Algorithm Of Image Feature Extraction And Classification Based On Nonnegative Matrix Factorization

The continuous upgrading of various types of data acquisition equipments has produced large amounts of data all the time,and the network has facilitated the rapid dissemination of these data.People obtain these data to take advantage of information that is useful to them.These data are often high-dimensional,so it is very necessary to research how to conduct dimensionality reduction so that effective low-dimensional information can be obtained from the data.As an effective method for data reduction,NMF approximates the input matrix as the product of two low-dimensional matrices,of which the elements are all nonnegative.In the field of data processing,nonnegativity is consistent with the cognition of the human brain and is of great interpretability.Meanwhile,factorization reduces the dimension of data,which can make the elements sparse and is conducive to data analysis.Non-negative matrix factorization plays a good role in image processing,but in the actual image analysis process,image data may be polluted or damaged by noise and the computational efficiency needs to be improved.So,based on the summary of the some improved non-negative matrix factorization methods,the dissertation proposes some improvements and solutions to existing problems,and the main research contents are as following:(1)A Sparse Corruption Non-Negative Matrix Factorization(SCNMF)method is proposed.The method takes both uncorrupted images and corrupted images into consideration,in which the uncorrupted matrix is represented as the product of the base matrix and the coefficient matrix,the corrupted matrix is approximated as the sum of two low-dimensional matrices multiplication and a noise matrix.The model simultaneously considers the two approximations and iterates the low-dimensional features of the data.In the experiments,extracted features are verified easy to distinguish,which is helpful for image classification.And for images under non-Gaussian noise or non-Poisson noise,the proposed method can reconstruct these images well.(2)A Label-Embedding Nonnegative Matrix Factorization(LENMF)is proposed.Traditional nonnegative matrix factorization methods do not take label information into consideration while factoring,while the label information of images is introduced in the proposed method.The proposed method separates a sparse noise matrix out of the input matrix,and the rest of the input matrix is represented as the product of two low-dimensional matrices.The feature matrix is limited by label a matrix so that samples with the same labels can be divided into the same subspace.As is demonstrated in the experiments,the proposed method overcomes the influence of image noise effectively and the base images are discriminative after iteration.(3)A Label-Embedding Online Nonnegative Matrix Factorization(LEONMF)is proposed.The subspace of traditional NMF methods is needed to be recalculated when new data is to be classified,which resulting in time-consuming and effort-consuming problems,so a new algorithm is proposed in the dissertation.Firstly,the base matrix is calculated by LENMF method to reduce the data.Then,the base matrix is adjusted and connected with the testing matrix to get the new matrix.Finally,the connected matrix is decomposed for a new weight matrix which is available for the classification the new data after adjustment.The base matrix does not need to be repeatedly calculated in this process and only the new data needs to be connected to the existing base matrix for calculation when new data appears.Experiments prove the effectiveness of the proposed method,and prove that the method is time-saving.