A Hyperspectral Target Detection Model And Algorithm Based On Structured Tensor Representation

Hyperspectral image has the characteristic of acquiring spectral and spatial information simultaneously,its rich spectral and spatial information provide the fundamental basis for identifying ground materials.Anomaly detection is practical and universal,and does not require the prior knowledge of the target,which based on the difference of background statistical characteristics.Therefore,hyperspectral anomaly target detection become a research hotspot both at home and abroad.Traditional target detection methods usually process and analyze hyperspectral data using vector or matrix form,which will destroy the original spatial structure features.Tensor representation is an effective tool for analyzing high-dimensional data.The high-order tensor structure can describe the multi-feature characteristic of hyperspectral images,and maintain the spatial spectrum information of the image better.In this paper,using low-rank representations tensor decomposition as the basic framework,we analyze the spatial-spectrum structure features of hyperspectral images in-depth.Hyperspectral image anomaly detection methods based on tensor structure representation is proposed.The main work of this paper includes:(1)We propose a hyperspectral image anomaly detection method based on low-rank and sparse representation tensor decomposition.This method decomposes the original hyperspectral tensor data into a background tensor and an anomaly tensor,the background is modeled using low-rank tensor decomposition,a sparse regularization is used to characterize the anomaly part.At the same time,we use the K-means clustering algorithm to classify the hyperspectral material pixels,and design a method to construct a spectral dimension dictionary factor matrix based on kernel spectral angle cosine.In view of the anomaly detection experiments both on real hyperspectral data sets and simulated data sets,our proposed method has a lower missing error rate and false alarm rate and achieves good anomaly detection results.(2)We propose a hyperspectral image anomaly detection based on mahalanobis distance regularized tensor decomposition.In this method,structured feature representation of hyperspectral data is achieved by Tucker decomposition.Then we construct the spatial dimension dictionary factor matrix to extract spatial information by using High Order Singular Value Decomposition(HOSVD),and map the hyperspectral tensor data to the low-dimensional subspace.In order to extract spectral information,spectral dimension factor matrix is constructed by using spectral dictionary construction method.For further increasing the effectiveness of anomaly detection,a mahalanobis distance regularization term is introduced to characterize the background statistics,in which the background and anomalies can be effectively separated.The experimental results show that the mahalanobis distance regularized tensor decomposition model proposed in this paper can effectively suppress background and highlight anomal targets,and the detection AUC value is more than 0.98.