Research On Image Subspace Clustering Method Based On Low-rank Tensor

With the rapid development and widespread popularity of computer network technology and multimedia technology,the form and quantity of information has exploded.Therefore,in machine learning,in order to fully characterize a research object,its data usually has high-dimensional characteristics.However,the noise and redundant information present in these data can cause training deviations and complex parameter for subsequent models.Therefore,preprocessing the specific data set and obtaining the appropriate feature subspace is an important data preprocessing method.In recent years,researchers have done a lot of work to improve the ability to express the original feature space.Based on the low rank tensor theory,this paper conducts the following research work on feature representation.Aiming at the clustering of image subspace,this paper proposes a novel low-rank tensor approximation method to better explore the underlying essential structure of tensor data and improve the clustering performance of features.The algorithm aims to obtain a low-dimensional kernel tensor representation by projecting the original data to the potential subspace.Compared with the previous low rank approximation methods,the low rank constraint is added to the projection matrix to solve the low rank kernel tensor.Our algorithm directly imposes a low rank constraint on the nuclear tensor.For the feature representation of image dataset,this paper proposes a feature learning method based on regularized low-rank tensor constraint,which aims to remove the noise and redundancy of the original feature space and improve the expression ability of the original feature space.The algorithm mainly consists of two parts: the low-rank tensor approximation of the original feature space and the feature selection of the graph regularization.The tensor expression of the data not only preserves the spatial structure information of the original data,but also reduces the computational complexity.The low-rank approximation is to remove the noise and redundant information by minimizing the reconstruction error,thereby exploring the potential information of the original features and mapping to the appropriate feature subspace.The feature selection algorithm based on graph regularization can preserve the local geometric structure information of the feature for ensuring the tightness between the same category of samples and the separability between different categories of samples.