Camera-independent learning and image quality assessment for super-resolution.

摘要

An increasing number of applications require high-resolution images in situations where the access to the sensor and the knowledge of its specifications are limited. In this thesis, the problem of blind super-resolution is addressed, here defined as the estimation of a high-resolution image from one or more low-resolution inputs, under the condition that the degradation model parameters are unknown. The assessment of super-resolved results, using objective measures of image quality, is also addressed.; Learning-based methods have been successfully applied to the single frame super-resolution problem in the past. However, sensor characteristics such as the Point Spread Function (PSF) must often be known. In this thesis, a learning-based approach is adapted to work without the knowledge of the PSF thus making the framework camera-independent. However, the goal is not only to super-resolve an image under this limitation, but also to provide an estimation of the best PSF, consisting of a theoretical model with one unknown parameter.; In particular, two extensions of a method performing belief propagation on a Markov Random Field are presented. The first method finds the best PSF parameter by performing a search for the minimum mean distance between training examples and patches from the input image. In the second method, the best PSF parameter and the super-resolution result are found simultaneously by providing a range of possible PSF parameters from which the super-resolution algorithm will choose from. For both methods, a first estimate is obtained through blind deconvolution and an uncertainty is calculated in order to restrict the search.; Both camera-independent adaptations are compared and analyzed in various experiments, and a set of key parameters are varied to determine their effect on both the super-resolution and the PSF parameter recovery results. The use of quality measures is thus essential to quantify the improvements obtained from the algorithms. A set of measures is chosen that represents different aspects of image quality: the signal fidelity, the perceptual quality and the localization and scale of the edges.; Results indicate that both methods improve similarity to the ground truth and can in general refine the initial PSF parameter estimate towards the true value. Furthermore, the similarity measure results show that the chosen learning-based framework consistently improves a measure designed for perceptual quality.