Pan-sharpening with multi-scale wavelet dictionary

In satellite image processing, pan-sharpening is the fusion process in which a low resolution (LR) multi-spectral (MS) image is sharpened using the corresponding high resolution (HR) panchromatic (Pan) image to obtain a HR MS image. In this paper we propose a novel pan-sharpening method which combines the ideas of classical wavelet-based pan-sharpening with recently developed dictionary learning (DL) methods. The HR MS image is generated using wavelet-based pan-sharpening, regulated by promoting sparsity with respect to a dictionary. The dictionary is obtained using DL on the multi-scale wavelet tree vectors of the image to be pan-sharpened. A significant advantage of our approach compared to most DL-based approaches is that it does not require a large database of images on which to train the dictionary. Experiments on degraded satellite images demonstrate that our method significantly reduces color distortions and wavelet artifacts compared to the state of the art.