Denoising of Multispectral Images via Nonlocal Groupwise Spectrum-PCA

Aram Danielyan, Alessandro Foi, Vladimir Katkovnik,
Karen Egiazarian
Tampere University of Technology, Finland
Download paper
Play (15min) Download: MP4 | MP3

We propose a new algorithm for multispectral image denoising. The algorithm is based on the state-of-the-art Block Matching 3-D filter. For each “reference” 3-D block of multispectral data (sub-array of pixels from spatial and spectral locations) we find similar 3-D blocks using block matching and group them together to form a set of 4-D groups of pixels in spatial (2-D), spectral (1-D) and “temporal matched” (1-D) directions. Each of these groups is transformed using 4-D separable transforms formed by a fixed 2-D transform in spatial coordinates, a fixed 1-D transform in “temporal” coordinate, and 1-D PCA transform in spectral coordinates. Denoising is performed by shrinking these 4-D spectral components, applying an inverse 4-D transform to obtain estimates for all 4-D blocks and aggregating all estimates together. The effectiveness of the proposed approach is demonstrated on the denoising of real images captured with multispectral camera.