A noise reduction algorithm of hyperspectral imagery using double-regularizing terms total variation

In the present paper, an effective total variation denoising algorithm is proposed based on hyperspectral imagery noise characteristics. The new algorithm generalizes the classical total variation denoising algorithm for two-dimensional images to a three-dimensional formulation. Considering the fact that the noise of hyperspectral imagery shows different characteristics in spatial domain and spectral domain respectively, the objective function of the proposed total variation algorithm is improved by utilizing double-regularizing terms (spatial term and spectral term) and separate regularization parameters respectively. Then, the new objective function is discretized via approximating the gradient of the regularizing terms by three orthogonal local differences, and further majorized by a convex quadratic function. Thus, noise in spatial and spectral domain could be removed independently by minimizing the majorizing function with a majorization-minimization (MM) based iteration. The performance of the proposed algorithm is experimented on a set of Hyperion imageries acquired in 2007. Experiment results show that, properly choosing the values of regularization parameters, the new algorithm has a similar improvement of signal-to-noise-ratio as minimum noise fraction (MNF) method and Savitzky-Golay filter, but a better performance in removing the indention and restoring the spectral absorption peaks.