Multiscale Tensor Dictionary Learning Approach for Multispectral Image Denoising

Taking advantage of different sensitivities of each component/material in an object to different bands, multispectral image (MSI) is obtained by shooting the object from multiple bands individually or simultaneously. Including these different complementary information with the space-time correlation, the MSI can describe the object more clearly and comprehensively. In practice, however, it is unavoidable that MSIs are corrupted by noise. In order to solve the denoising problem of MSIs, a framework is proposed to suppress noise by learning multiscale sparse representations of MSIs with an overcomplete tensor dictionary. In our method, the tensor patches are extracted from an image tensor, and a tensor-based dictionary is trained using a special tensor decomposition, in which each atom is a rank-one tensor. The so-called multiscale learned representation is obtained based on an efficient quadtree decomposition of the trained tensor dictionary. Experimental results on numerical simulations and real MSIs demonstrate that multiscale tensor dictionary gets better indexes in terms of PSNR and SSIM compared with single-scale tensor dictionary and other related competing methods. At the same time, from the perspective of visual quality, our method restores more image details than other methods.