Research Of Image Feature Extraction Algorithm Based On Latent Low-rank Representation

The rapid development of big data and information technology has made the information obtained from reality larger and more complex,.and its characteristics show a trend of high dimension and unstructured.Moreover,high-dimensional data is often accompanied by a large amount of noise and redundant information,which not only leads to a waste of computing resources and memory consumption,but also leads to a decline in algorithm performance.Therefore,how to robustly extract effective features from high-dimensional data and apply them to tasks such as pattern recognition has become an important research topic.In response to this issue,this article studies high-dimensional manifold data,revealing the global structure of high-dimensional manifold data through latent low-rank representation,enhancing learning ability and robustness against under-sampling and severely damaged data,and proposing two feature extraction optimization models based on this.The main research content and contributions of the paper are summarized as follows:（1）A latent low rank sparse projection feature extraction method based on dual neighborhood preservation and feature selection is proposed to address the drawback of extracting the main and significant features separately from latent low-rank representation,which cannot maintain the local manifold geometry and explain the importance of features.This algorithm comprehensively considers the interaction between row space and column space information,integrates data reconstruction and low-rank representation into one term,and replaces the previous salient feature extraction matrix with two matrices.This enables the model to not only flexibly extract features but also maintain the global structure of the data.At the same time,in order to mine the manifold geometry of the data feature space,neighborhood preserving regularization is applied to both the low rank self-representation matrix and low dimensional data.Next,apply_2,1norm to the projection matrix,_2,1norm sparse constraints are used for feature selection,making the extracted features more interpretable.In addition,in order to further improve the robustness to noise,_2,1norm is introduced.The_2,1norm is used to constrain the noise component.（2）A discriminative low-rank preserving embedding feature extraction method based on non-negative sparse graph regularization reconstruction is proposed to address the problems of complex algorithm models,poor quality of discriminative feature distribution,and low interpretability in the past.By imposing the nearest neighbor graph regularization constraint on the data reconstruction error based on the latent low-rank representation,the algorithm uses only one term to retain the global and local structure of the data at the same time,which greatly reduces the complexity of the model compared with the previous model using multiple regularization terms and balance parameters.Subsequently,the relationship between the low rank representation coefficient of the data and the discriminant similarity is modeled,capturing the block diagonal structure of the low rank self-representation matrix of the data,and enhancing the interpretability of the low rank self-representation.Finally,a sparse term of_2,1norm is applied to the projection matrix to make the features more compact and discriminative.Based on the above model,this article designs corresponding optimization algorithms,using various methods including alternating direction multiplier method,singular value decomposition,and singular value threshold method.In addition,this article also conducted a theoretical analysis on the time complexity and parameter sensitivity of the model.Extensive experiments on multiple publicly available datasets have shown that the proposed method is more effective and robust than the comparison algorithms.