Super-resolution of Multispectral images
The topic of Super Resolution (SR) or Resolution Enhancement (RE) as used here refers the process of obtaining a High Resolution (HR) image or a sequence of HR images from a set of Low Resolution (LR) observations. It first appeared in the early 1980s and since then the topic has been active, with some of the early good results appearing in the 1990s. The last five or so years, however, have witnessed an enormous resurgence in SR activity. This assessment is supported by the number of the recent papers on the topic but also with the on-going R&D projects on super-resolution in various segments of the industry and various research laboratories. In parallel with the activities in the signal processing community increased activity was observed recently in the area of applying Bayesian modeling and inference to the SR task as well as in pattern recognition and computer vision communities in applying learning techniques to the SR problem. We believe that the SR/RE area has matured enough to develop a body of knowledge that can now start to provide useful and practical solutions to challenging real problems. This indeed provides an explanation for the renewed interest in it.
SR techniques are being applied to multispectral images when several multispectral cubes of the same scene have been observed and such cubes are shifted at the sub-pixel level. The first work on SR was aimed at improving the resolution of Landsat images. When dealing with multispectral images, the term SR (or pansharpening) is also used to define the problem of combining HR images which are observed with panchromatic sensors with LR images observed at a given wavelength.
In this presentation, in order to tackle the SR reconstruction of multispectral images we follow the hierarchical Bayesian framework to incorporate prior knowledge on the expected characteristics of the multispectral images, to model the observation process of both panchromatic and LR multispectral images, and also to include information on the unknown parameters in the model in the form of hyperprior distributions. Then, by applying variational methods to approximate probability distributions we estimate the parameters of the hyperprior distributions on the unknown parameters together with the unknown parameters, and the HR multispectral image.