Research On Hyperspectral Remote Sensing Imagery Target Detection Based On Sparse Representation

Hyperspectral remote sensing imaging is different from conventional optical sensing imaging.Hyperspectral remote sensing imaging represents thousands of band information by a cube image which contains image information and spectral information.Hyperspectral remote sensing imaging integrates the spatial and spectral information of into a form of a three-dimensional cube image.Compared with multispectral images,hyperspectral remote sensing images have more spectral band information.The use of subtle differences between the spectral information of hundreds of objects can achieve target recognition and feature classification.Target detection can be divided into supervised target detection and unsupervised target detection according to whether prior information of terrain is needed or not.In this paper,the basic principle of hyperspectral remote sensing imaging technology is combined with the sparse representation model,and the original signal is linearly expanded in the over-completed dictionary.The sparse coefficients are solved by orthogonal matching pursuit algorithm and the original signal can be reconstructed by the over-completed dictionary and sparse coefficients.Based on the unique spectral reflection curves of different features,the discriminant function is designed and the appropriate threshold is selected to distinguish the target from the background.The main contents of this paper are as follows:First of all,the over-complete dictionary in the original target detection algorithm based on sparse representation contains a small number and a few kinds of training samples,which will have an effect on the result of target detection.Aiming at solving the shortcomings of traditional dictionary target detection algorithm,this paper increases the number of target training samples in the over-completed dictionary by proliferating the original dictionary,and improves the accuracy of target detection.The effectiveness of this algorithm has been verified by comparison with the original sparse representation algorithm.Secondly,the traditional target detection algorithm sparse representation-based target takes too long time to solve sparse vectors and only consider the spectral information of hyperspectral images in the detection process,and also does not consider the issue of spatial information.In this paper,by changing the idea of sparse vector,the target and the background sparse vector are solved step by step,and the original model is improved in the detection process.The spatial information is added to improve the target detection accuracy and speed up the target detection time.Finally,hyperspectral imagery mixed pixels target detection based on dictionary reconstruction is proposed.In the detection process,the structure of the dictionary is obtained directly from the hyperspectral image and there are uncertainties.Meanwhile,the traditional sparse representation algorithm can not directly detect the mixed pixels.In order to solve the following problems,this paper proposes hyperspectral imagery mixed pixels target detection based on dictionary reconstruction,using an unsupervised method to complete the construction of dictionary to ensure that the dictionary contains some spectral information of target pixels.The detection of mixed pixels in hyperspectral images is realized by introducing binary hypothesis model.Experiments have been carried out on hyperspectral images,which reveals that the method proposed shows an outstanding detection performance.