Sparse Representation With Spatial Information For Target Detection In Hyperspectral Images

With higher and higher spectral resolution,hyperspectral images(HSI)include plenty of spectral bands,which could convey ample spectral information and distinguish minute spectral differences between similar land features.Therefore,they could provide unique superiority for target detection.In recent years,the algorithms in target detection have made some research progress,but there are still some problems to be solved,including how to exploit spatial information more effectively.This paper mainly introduces evaluation criteria for target detection and the classification of detectors.Representative detectors are described in four categories: detectors based on spectral matching,spectral decomposition,binary hypothesis,and sparse representation.According to defects and merits of detectors based on sparse representation,two detectors based on sparse representation with spatial information are proposed.Firstly,although a few sparsity-based detectors exploit neighboring information,and perform spatial-spectral joint detection.However,it is often assumed that the central pixel to be detected and its neighboring neighborhood pixels belong to the same object type.The neighboring pixels participate in the sparse reconstruction of the central pixel according to the same weight.This assumption is often different from the actual situation in the heterogeneous regions of the image,leading to the poor model generalization ability.To alleviate this problem,considering the similarity between different neighborhood pixels and central pixels,this paper proposes the spatially adaptive sparse representation for target detection.On the other hand,the background dictionary is often constructed by using a dual concentric window in sparsity-based detectors.Inevitably,the target pixel atoms may fall into the background dictionary,which will reduce the separability between target and background.At the same time,the over-completed dictionaries are difficult to obtain.Besides,the factors such as imaging environment and complex distribution of features often lead to the fact that a small number of background pixels cannot reconstruct the spectrum of the central pixel well.In order to alleviate this situation,a binary hypothesis algorithm based on a discriminative subspace and adaptive dictionary is proposed by constructing a discriminative subspace and combining the collaborative representation with spatial information to further improve the target detection effect.The algorithms were tested on real HSIs,and it was found that the proposed algorithms in this paper overcome the defects of detectors based on sparse representation to some extent and show an outstanding detection performance.