Abstract: The acquisition of hyperspectral images (HSI) is often interfered by multiple types of noise,which will directly affect accuracy in the subsequent applications.Therefore,HSI denoising is a very important pretreatment process.The low-rank representation (LRR) model can well satisfy the spectral properties of HSI.However,the choice of dictionary under this framework is particularly significant,which is still an open question at present.Meanwhile,the typical method can’t satisfy the requirement well by only considering the local correlation of the image,and the non-local similarity is equally of significance.Based on LRR,a new method of HSI denoising was proposed.Firstly,the type of noise is considered comprehensively and the dictionary with more comprehensive discrimination ability is selected.Secondly,on the premise of block processing,non-local similar information is introduced through clustering,and similar blocks are combined for LRR framework.The experimental results on the simulated In-dian Pines and real EO-1 Hyperion data set show that the proposed method performs better than the state-of-art HSI denosing methods both in the visual effect of the image and the quantitative evaluation index of the simulated data set.

Key words: Denoising, Dictionary selection, Hyperspectral images (HSI), Low-rank representation (LRR), Non-local similarity