Research On Dimension Reduction Technology And Its Application In Image Classification And Recognition

Dimensionality reduction is of great importance in many applications,such as pattern recognition,text categorization and computer vision where the dimensionality of data is often very high.It can reduce the computational complexity and discover the intrinsic manifold structure of high-dimensional data.Principal component analysis is perhaps one of the most famous dimensionality reduction technique,but it still has some shortcomings and defects,in this paper,the main work of this kind of problem is as follows:Firstly,this paper proposes an L2,1-norm-based sparse principal component analysis with trace norm regularized term.The L2,1 norm containing the mean of the sample to be optimized is used as the reconstruction error,while the regularization term uses the trace norm dependent on the sample matrix.In the process of calculating the projection matrix,the optimal mean is adjusted for each iteration.The algorithm is robust and the method is adaptive to the relevant structures present in the sample data.Secondly,combining the self-step learning method with the principal component analysis algorithm,a principal component analysis method based on self-step learning is proposed.Through the self-step regularizer,the algorithm can automatically select the sample from "simple" to "complex",through the L2,1 norm regularization penalty term to obtain the sparse projection matrix,the main features of the sample are extracted.Furthermore,a principal component analysis method combining joint low rank preservation and optimal mean is proposed.For the general principal component analysis method,only the overall geometry of the data can be preserved,and the shortcomings of the intrinsic manifold structure of the data cannot be well revealed.The global geometric information and the identification structure of the sample self-represented coefficient weight matrix can be retained in the low-dimensional embedded subspace,and the inherent manifold structure of the sample can be revealed while learning the optimal mean and the optimal projection matrix.Finally,a local linear embedded principal component analysis based on cooperative representation is proposed.The L2 graph based on the collaborative relationship is established,and the neighbors of the sample are determined.LLE and the improved PCA are integrated into the same model.The disadvantage of setting the number of samples nearest neighbors subjectively is avoided,while realizing feature extraction jointly.Experiments on some public data sets show that the proposed algorithms are feasible and effective,and has certain application value in image classification and recognition.