Hyperspectral Image Denoising using Dictionary Learning

Hyperspectral images are corrupted by noise during their acquisition. In this work, we propose to efficiently denoise hyperspectral images under two assumptions: (i) noiseless hyperspectral images in matrix form are low-rank, and (ii) image patches are sparse in a proper representation domain defined through a dictionary. These two assumptions have already led to state-of-the-art denoising methods using fixed Wavelet transforms. We propose to rather learn the dictionary from hyperspectral images, a task commonly known as dictionary learning. We show that the dictionary learning approach is more efficient to denoise hyperspectral images than state-of-the-art methods with fixed dictionaries, at the cost of a larger computation time.