Research On Hyperspectral Image Target Detection Algorithm Based On Sparse Representation

Hyperspectral image target detection is one of the important research directions of hyperspectral remote sensing technology.This type of algorithm can use the rich spectral information of hyperspectral images to identify target objects,and has unique advantages over traditional target detection algorithms.Among many hyperspectral target detection algorithms,the rapid development of sparse representation methods has emerged in many algorithms.However,in actual detection,due to the limitation of the quality of sparse representation training samples and the neglect of spatial information,it is difficult for sparse representation algorithms to achieve better detection results.This paper proposes a target detection algorithm for the sample quality and joint spatial and spectral information in the sparse representation algorithm.In this paper,the following researches on sparse representation algorithms are conducted:Because the sparse representation algorithm lacks the use of spatial information,a sparse representation method based on weighted joint k-nearest neighbor and multi-task learning is proposed.In terms of spectral information,the information connection of adjacent bands is considered on the basis of sparse representation,combined with multitask learning ideas,a multi-task learning sparse representation model is constructed.In terms of spatial information,the spatial pixel information in the neighborhood of the test pixel is considered,and the weighted Euclidean distance is used to highlight the different contributions between pixels.Finally,a detector is designed in conjunction with space-spectrum information to perform more stable and accurate target detection tasks.Compared with the algorithm that does not use spatial information,the proposed algorithm has a higher detection effect.The AUC values on the AVIRIS and Texas Coast datasets are 99.12% and 97.33%,respectively,which proves that the algorithm can effectively detect the target.The quality of the sparse dictionary of sparse representation is often poor,and a method of constructing a sparse dictionary based on superpixel segmentation is proposed.First,the super-pixel segmentation algorithm is used to segment the reduceddimensional hyperspectral image.Then use the prior information of the target to calculate the correlation coefficient in the local super pixel block,and finally use the correlation coefficient to select high-quality target training samples to construct a highquality sparse dictionary to improve the target detection effect of the sparse representation.Superpixel segmentation algorithm,through spectral and spatial information,the pixels with similar characteristics are collected into super pixel blocks,which have local spatial information,and the correlation coefficient further obtains high-quality target samples,builds high-quality sparse dictionary,and obtains better training effect.The experimental results show that the AUC values of the proposed algorithm in the AVIRIS and Texas Coast datasets are 99.35% and 98.79%,respectively,which have a higher detection effect than other algorithms.Aiming at the problem that the quality of sparse training samples is degraded by the phenomenon of spectral variation,a training sample optimization method based on sparse representation and spectral angle is proposed.First,extract the target pixels and candidate pixels from the image,and construct a sparse representation,Then count the sparse representations of the high-frequency dictionary atoms in the result,and construct the optimized target sample through the high-frequency dictionary atoms.Finally,high-quality training samples are obtained through the spectral angular distance.The proposed algorithm obtains the most representative target sample set based on the characteristics of sparse representation,and the optimized target sample performance obtained thereby reduces the influence of spectral variation and improves detection performance.Through experimental comparison,due to the optimized processing of the target prior pixels of the hyperspectral image,the algorithm has a good detection effect in different data sets,especially in the AVIRIS data set,the AUC index reaches 99.78%,which proves the effectiveness of the algorithm.