Regularized super-resolution of multi-view images.

摘要

Super-resolution is the process of obtaining either a higher resolution still image or a sequence of higher resolution images from a corresponding sequence of low resolution images of a particular scene. It extends the performance limits of single image interpolation by leveraging the unique information present in the multiple albeit slightly different images. These multi-image techniques seek to recover frequency content beyond that present in any of the individual observed images and are hence termed 'super-resolution' algorithms. In its simplest form, a super-resolution algorithm aims to align the mutually shifted low resolution images on a higher resolution grid. The alignment process requires precise knowledge of the displacement occurring in the scene, which is estimated using the low resolution images. This allows a formulation of multiple observed data constraints that can be used together with knowledge about the imaging process to estimate the high resolution image. Although super-resolution algorithms have been shown to perform well in synthetic scenarios, many of the modeling assumptions break down in real world imaging conditions. Super-resolution performance is then heavily dependent on how the forward imaging model is constructed, which is a recurring theme of this thesis. We constrain ourselves to a specific imaging device and examine the effects of super-resolution when the characteristics of the camera are uniquely identified. Displacement estimation has been identified as a major factor in the performance of super-resolution and the choice of displacement models for different scenes is examined in the thesis. The estimation of the high resolution image is carried out using regularization-based methods (both algebraic and stochastic). The thesis also addresses artifacts arising from inaccurate displacement estimates either due to inconsistent displacement models and/or occlusions occurring in the scene. A complete system is built in this thesis and the results obtained show significant improvement over single image bi-cubic interpolation.