Multi-directional Total Variation Hyperspectral Image Destriping based on Tensor Decomposition and Group Sparsity

Most of the traditional destriping work is based on matrix domain (band-by-band) processing hyperspectral images, although satisfactory results can be obtained, some important prior conditions are still ignored. In fact, a hyperspectral remote sensing data can be regarded as a third-order tensor, and the space-spectrum Total Variation based on the tensor has been widely used in destriping work. However, the processing of the strip part rarely involves tensor operations. Driven by the high correlation between hyperspectral data in different spectral bands, in this paper, we proposed a multi-directional total variation destriping algorithm based on tensor decomposition and group sparsity. For the target image, a multi-directional full variational regularization constraint is used to ensure piecewise smoothness while preserving continuity between adjacent spectrums. For stripe noise, we adopted low-rank Dicker decomposition and weighted L21 norm joint constraints. Low-rank Dicker decomposition treats stripe noise as a tensor data, which can better describe global correlation; Weighted L21 norm is used to describe group sparsity of stripe noise. A series of experimental results show that our algorithm can effectively remove all kinds of stripe noise while better retaining the detailed information of the target image.