Super-Resolution Using Adaptive Wiener Filters
This chapter will describe an approach to super-resolution (SR) based on adaptive Wiener filters. In this approach, multiple frames are registered relative to a common grid. Output SR pixel values are estimated using a type of adaptive Wiener filter that forms a weighted sum of neighboring observed pixels on the registered high-resolution grid. The filter weights are selected to minimize the mean squared error based on statistical correlation models or empirical training data. This estimation step simultaneously serves both a nonuniform interpolation function and a restoration function. This SR approach is most appropriate when the point spread function model and the motion model commute. If the filter weights can be precomputed, this approach can lead to very fast implementations. Note that the weights adapt to different spatial positions of the observed samples and can also be designed to adapt to local structure in the intensity data (e.g., edges, lines, and flat regions). We consider adaption based on vector quantization and local variance.