A Mixed-Norm Fidelity Model for Hyperspectral Image Denoising under Gaussian-Impulse Noise

In this paper, we evaluate an optimization problem to denoise high-dimensional hyperspectral data. We formulate the problem as a variational technique for simultaneously handling mixed Gaussian and impulse noise which is the characteristic of hyperspectral data. Vectorization of data and subsequent application of Total Variation as a regularization term exploit correlation along spatial and spectral dimension. We then use a mixed data fidelity model to simultaneously handle the effect of mixed Gaussian-impulse noise. The resultant optimization problem is solved using gradient descent. Experimental results on synthetic and real hyperspectral data support our argument of using a mixed-norm fidelity term.