A Tutorial on Multi-frame Computational Super-resolution using Statistical Methods

The term super-resolution is used to describe methods aimed at recovering detail information of an imaged scene that would otherwise be lost during the normal imaging process. Here we discuss only digital cameras that process light reflected by the objects constituting a scene. A digital camera, consisting of optics a digital imaging sensor and signal processing hardware is able to capture a digital image of a scene. For digital cameras, resolution has to do with both the number of imaging elements per unit sensor area – which translates into the number of pixels per unit image area – as well as the information content of the digital image. Intuitively speaking, increasing the information content is directly related to increasing the number of pixels per unit area in the digital image such as to properly render the enlarged information content according to the Shannon sampling theorem. This tutorial is concerned with increasing the level of detail (i.e., information content) in a digital image by means of statistical processing algorithms starting from a set of several (usually) alias-afflicted images of the same scene, acquired from slightly different positions. Such techniques fall in the category of multi-frame computational super-resolution. We start be describing the maximum likelihood (ML) solution to this problem and then show how a maximum a posteriori (MAP) approach can improve upon the ML solution. We will discuss several solution strategies relying on such principles and point to their advantages and disadvantages. We conclude with a short overview of alternative super-resolution approaches.