Hyperspectral Image Denoising And Classification

Denoising and classification are the two most basic applications of hyperspectral images.During the acquisition of hyperspectral images,noise is unavoidably introduced due to photon effects or sensors.Therefore,denoising of hyperspectral images is particularly important,and low rank-based methods are one of the effective denoising methods.Based on the deep analysis of the low-rank algorithm,this paper proposes a new low-rank denoising algorithm.For the classification of hyperspectral images,dictionary learning is an effective classification method.Based on the deep analysis of dictionary learning,this paper proposes a new dictionary learning algorithm.The main contents include:(1)Considering the complexity of hyperspectral images,the traditional segmentation method can not handle the samples near the boundary of the image category.The superpixel segmentation method can make the same block contain the same type of samples.Secondly,in order to make better use of the target rank prior knowledge,the PSSV model is used to denoise each sub-block.Then a denoising model based on super-pixel segmentation and PSSV is proposed,namely SS-PSSV.Experimental results in three common hyperspectral databases show that the proposed method can achieve better denoising results compared to common denoising models.(2)Considering the existence of abnormal values of hyperspectral images,in order to be able to characterize the nonlinear features of hyperspectral images and suppress the abnormal values,this paper improves the dictionary learning model LC-KSVD and proposes a new one dictionary learning model,namely ELC-KSVD.ELC-KSVD first maps the hyperspectral image to Euler space through Euler expressions,and then trains the dictionary in Euler space.This not only weakens the influence of noise,but also does not increase the sample dimension in Euler space.For the model optimization problem of ELC-KSVD,this paper proposes a new solution method,namely Euler sparse coding.Experimental results in three common hyperspectral databases show that the proposed method ELC-KSVD has better classification performance.