Research On Hyperspectral Image Target Detection Based On Sparse Representation

With the development of remote sensing technology,hyperspectral remote sensing technology has been spawned.Hyperspectral remote sensing technology is an imaging spectral technology which can combine spectral information and two-dimensional images,so hyperspectral images are a fusion of spatial information and spectral information.Different objects have different spectral characteristics,and even the same objects in different states have different spectral characteristics,which is the basis for hyperspectral images to be able to perform target detection.In this thesis,the data characteristics of hyperspectral images are combined with sparse representation model,and investigate the target detection model based on sparse representation of hyperspectral images with the provision of a priori target spectral information.The main research contents are as follows:1.There are still some problems to be solved in the target detector based on sparse representation.In hyperspectral images,the proportion of target pixels is usually small relative to background pixels,that is,the existing target detection algorithms based on sparse representation model often select target training samples from the global image,resulting in insufficient target training samples and unbalanced data volume compared with the background training samples,so it is difficult to achieve the desired detection performance.To tackle these problems,a sparse representation target detection algorithm based on oversampling and a binary hypothesis sparse representation target detection algorithm based on oversampling are proposed.Both algorithms use a new target dictionary construction method to extract the initial target training samples by constrained energy minimization method,and then expand the target training samples by using synthetic minority oversampling technique.Experimental results show that the detection performance of the proposed algorithms is improved and the problem of insufficient samples is effectively solved.2.Aiming at the problem that the construction process of background dictionary in the sparse representation model is disturbed by target elements and the target and background sparse vectors affect each other in the process of solving sparse vectors,thus affecting the detection performance,a binary-class sparse representation hyperspectral image target detection algorithm based on background dictionary construction is proposed.The method utilizes low-rank and sparse matrix theory,and introducing the target dictionary as a priori information to enhance the target and background separation capability.A pure background dictionary is constructed using decomposed low-rank background hyperspectral image,and a binary-class sparse representation model is used as the target detector.The experimental results show that the proposed algorithm has better target detection performance.