Fig. 2.

Block diagram of the signal recovery procedure. Local information within ball-shaped sliding patches is extracted by $\mathbf{P}$. Random signal plus noise matrices $\mathbf{Y}$ and covariance tensors ${\mathbf{\Lambda }}_{\mathbf{Y}}$ are constructed for each of the patches $[q\text{'}]$. SVD of these matrices is performed with singular vectors in $\mathbf{V}$ and ${\tilde{\mathbf{V}}}^H$ and singular values in ${\mathit{\eta }}_{\mathbf{Y}}$. In parallel, noise properties are modelled by a spectral distribution $f\left({\mathit{\eta }}_{\mathbf{W}}\right)$, which allows optimally shrunk singular values, ${\hat{\mathit{\eta }}}_{\mathbf{X}}$, to be computed. Matrix estimates $\hat{\mathbf{X}}$ are synthesized using the shrunk singular values and overlapped patch information is assembled onto the signal estimates $\widehat{\tilde{\mathbf{x}}}$.