A Study On The Spatial-spectral Feature Description Algorithm Of Hyperspectral Image Based On Tensor

Hyperspectral images have many bands,large amounts of data,and high redundancy,so how to efficiently describe the characteristics of hyperspectral images becomes one of the hotspots in current hyperspectral image processing and analysis.Traditional hyperspectral image feature description methods are mainly based on the analysis of spectral curves.However,the problems of “same object with different spectrum” and “same spectrum with foreign object” in hyperspectral images reduce the accuracy and robustness of traditional feature description algorithms.For this reason,the spatial-spectral feature description algorithm of hyperspectral images has attracted extensive attention from researchers at home and abroad in recent years.This paper focuses on the spatial-spectral feature description algorithm of hyperspectral images.In order to make full use of the spatial and spectral information of hyperspectral images,this paper uses tensor mathematics to establish the spatial-spectral tensor model of hyperspectral images.The model regards the hyperspectral image as an anisotropic data cube,without destroying the independence of spectral domain and not ignoring the spatial features of hyperspectral images.Based on this model,we propose two spatial-spectral feature description algorithms for hyperspectral images:(1)This paper presents a SIFT description algorithm based on tensor gradient in hyperspectral images,which is abbreviated as TGSIFT description algorithm.Firstly,the tensor model of hyperspectral image is used to define the tensor gradient.Then,the tensor gradient of each pixel in the neighborhood of the center point is calculated.The four-quadrant antitangent trigonometric function which maps the tensor amplitude angles is used to increase the amount of information of descriptor.Finally,a statistical histogram of tensor gradient is created and the statistical histogram of the tensor gradient is converted into a one-dimensional descriptor as a description of the center point.In order to verify the validity of the TGSIFT algorithm,the TGSIFT description algorithm is used in the matching experiment of hyperspectral images.The experimental results show that the TGSIFT description algorithm has good robustness under the changes of scale,illumination,noise,rotation,affine,etc.Inthis paper,the TGSIFT description algorithm is also used in the classification experiments of hyperspectral images,the experimental results show that TGSIFT description algorithm can fully explore the rich information contained in the spatial and spectral domains of hyperspectral images,and can significantly improve the classification accuracy of hyperspectral images.(2)A tensor-based local binary pattern description algorithm for hyperspectral imagery is proposed,which is abbreviated as TSSLBP description algorithm.Firstly,the tensor model of hyperspectral image is decomposed by tensor to obtain more refined and useful information.Then the joint distribution function of the hyperspectral image in the spatial-spectral domain is studied.We design the local binary pattern coding method in spatial-spectral domain for hyperspectral images,including the binarization coding on the mode 3 fiber and the local binary coding method on the front slice.Finally,the hyperspectral image joint local binary pattern coding descriptor is formed by the statistical histogram.The TSSLBP description algorithm capture information both in the spatial domain and the spectral domain,and can enrich the spatial domain and spectral domain texture information of the hyperspectral images.In order to verify the validity of the TSSLBP description algorithm,the TSSLBP description algorithm is used in the classification experiment of hyperspectral images.The experimental results show that the TSSLBP description algorithm can make full use of spatial information and spectral information,and can significantly improve the classification of hyperspectral images.In a word,given the significance of hyperspectral image space spectral feature description algorithm for image classification,target detection and other fields,this paper makes a tentative exploration of hyperspectral image description algorithm using tensor mathematics theory,and proposes two spatial domain feature descriptions.Experiments on real data sets and public data sets,verify the effectiveness of the algorithm.The work of this paper provides new ideas for theories and methods of hyperspectral image feature description algorithms and even hyperspectral image processing and analysis.