Research On De-striping Method Of Hyperspectral Remote Sensing Images

Hyperspectral remote sensing images(HSI)normally contains abundant spatial and spectral information,which have been widely used in many civil and military fields.Due to the restriction of imaging technology,the stripes noise will be introduced into the collected HSI data during imaging process.Thus,de-striping is a key step of subsequent hyperspectral images processing.This thesis studies the problem of hyperspectral remote sensing images de-striping,and proposes novel de-striping algorithms:(1)The traditional image decomposition theory is to decompose a degraded hyperspectral images into three part: clear images,stripes,random noise.In order to further utilize the a priori knowledge of clear images,in this thesis,a clear image is divided into three components(jump component,transition component,and gently component)by a pixel manner.Different from the traditional destriping model,which uses a uniform regular term to constrain the target image,three different regularization operators will be adopted adaptively for different components in this thesis.For the stripes,non-convex non-smooth iteratively reweighted nuclear norm is used to depict its low rank property,which can improve the robustness of the traditional nuclear norm.It is worth mentioning that the model proposed in this thesis can be regarded as a generalized model of the traditional image decomposition model,which can be degraded to the traditional image decomposition de-striping model under certain circumstances.(2)The traditional de-striping work is often based on matrix operations to process hyperspectral images band by band,completely discards the spectral information of hyperspectral images,which may destroy the spectral continuity of images.Thus,a novel multi-directional total variation destriping algorithm based on tensor decomposition and column sparsity is proposed in this thesis.First,multi-direction total variation regularization term is used to constraint the piecewise smoothness of clear images in both spatial and spectral space;For stripes,low-rank Tucker decomposition and weighted L21 norm joint constraints are used to ensure its low-rank property and column sparsity.Finally,a non-convex data fitting term is introduced to ensure the similarity between the clear image and the degraded image.A series of simulation experiments and real experiment results demonstrate that the algorithms proposed in this thesis are effective and efficient for hyperspectral images de-striping.