Research On MPCA Dimension Reduction Algorithm For Hyperspectral Image

Hyperspectral Image(HSI),which adds spectral dimension to ordinary two-dimensional images,has the advantages of spectral continuity and map integration,which can provide more information for the research of computer vision and other aspects,and has been successfully applied in many different fields.With the increasing number and complexity of data,higher requirements are put forward for the quality,storage space and transmission speed of hyperspectral images in real life.Therefore,in order to solve the above problems,dimensionality reduction of hyperspectral images is very necessary.This paper mainly does the following work:For hyperspectral images,the traditional way of dimension reduction is not considering the spatial information between different wave band,the method of typical mainly is the principal component analysis,the PCA is based on the vector for feature extraction,destroys the original hyperspectral data of spatial structure and data of the relationship between different spatial dimensions.Therefore,this paper introduced the tensor of hyperspectral image building,at the same time considering the spectral dimension.The spectral dimension and the spatial dimension between bands are used to construct the third-order tensor model,and the data is compressed by MPCA based on the tensor data.The MPCA projection matrix of each mode in the process of dimension reduction in the reserve of information is the same,but different tensor under different module contains the amount of information is different.Therefore,in view of the uncertainty of the dimension of each order projection matrix in MPCA,this paper studies the adaptive determination algorithm of the dimension of projection matrix in MPCA based on multi-objective optimization.Through experimental comparison,the results show the effectiveness of the model and algorithm proposed in this paper.Considering that the influence of eigenvalues corresponding to eigenvectors of different sizes in the projection matrix on dimension reduction was not considered in the process of MPCA dimension reduction.Therefore,this paper first proposed a kind of weight matrix multi-linear principal component analysis,to build a high-dimensional multi-objective model,through the high-dimensional multi-objective optimization algorithm to solve the optimal weight matrix,improve the dimension reduction effect.And according to the actual problems of this paper,the high-dimensional multi-objective optimization algorithm is improved.Experimental results show that the improved MPCA algorithm based on weight matrix proposed in this paper has better dimension reduction effect.Due to traditional MPCA is in the process of dimension reduction,error are assumed to be Gaussian distribution,but the noise of the actual data is often irregular,especially when the data may also exist in the abnormal points,based on the Gaussian distribution of principal components can lead to serious bias MPCA find abnormal points,thus deviate from the data of the real principal component.Therefore,this paper mainly studies the algorithm of outlier detection,proposes an adaptive DBSCAN outlier detection algorithm,builds a multi-objective model,and solves it through multi-objective optimization algorithm.The experiment proves the effectiveness of the method.