Hyperspectral restoration based on total variation regularized low rank decomposition in spectral difference space

This paper proposes a novel mixed noise removal method based on total variation regularized low rank decomposition in the spectral difference space (termed TVLRSDS) for hyperspectral imagery (HSI). Spectral difference transform has been demonstrated to be able to change the structure of noise (especially for the structured sparse noise, e.g., stripes or deadlines) in the original HSI, thus enabling low rank tools to effectively remove the mixed noise instead of treating it as one of the low rank components. In addition, as the fact that spectra in an HSI lie in a low dimensional subspace, and the adjacent pixels are highly correlative, it inspires us to simultaneously utilize the nuclear norm to exploit the global low rankness, and employ the total variation to include the local piecewise smoothness in the spectral difference space for mixed noise removal of HSI. The proposed model with all convex terms could be easily solved by alternating direction methods of multipliers (ADMM). The experimental results demonstrate the effectiveness of the proposed method.