Research On Method Of Compressed Sensing Reconstruction Based On K-means Clustering For Hyperspectral Images

The massive data of hyperspectral images makes the system cost and implementation difficulty of subsequent processing increase dramatically.People trying to find a way to approximate the original high-dimensional data with as few compressed data as possible on the premise of distortionfree reconstruction.Aiming at this goal,using the inter-band image correlation of hyperspectral image,the structural similarity is introduced into the clustering strategy of hyperspectral image as a distance evaluation index,and a compressed sensing reconstruction method based on K-means clustering is established for hyperspectral images.The method divides band images of hyperspectral image into reference-band image group and normal-band image group using K-means clustering algorithm,and then reconstruct the reference-band image using smooth L0 norm(SL0)algorithm based on compressed sensing in the wavelet transform domain.The normal-band image is reconstructed using an improved prediction model.Simulation results show that the reconstruction accuracy is significantly improved compared with the method before improvement.In addition,a similar method is developed by replacing the wavelet domain as curvelet domain and the fast iterative shrinkage thresholding algorithm with SL0 algorithm,the simulation results of which show that the reconstruction efficiency of this method is improved compared with the aforementioned methods.The new grouping strategy reconstruction method partially resolves the data redundancy problem often encountered by hyperspectral remote sensing technology.