Abstract
Image denoising is a fundamental process in image processing, its objective is to remove noise of the acquired data. During the years several methods for image reconstruction have been developed and as of today variational models are some of the best methods. These models are based on the minimization of a particular energy functional which balances two terms, one related to data fidelity and one to regularization. The result of the reconstruction is highly dependent on this last factor. Unfortunately, multiple numerical experiments have showed how the reconstructions obtained via the most known regularization terms still contain artifacts. In this thesis is presented a new strategy of supervised learning intended to achieve higher quality results based on the training of an operator in the regularization term. This technique is referred to as bi-level training scheme. Moreover, new variants of the scheme aimed to reduce a few issues will be presented. Lastly, a comparison between the reconstruction methods will be provided, to highlight the pros and cons of each strategy. Every statement will be backed up by the results of the numerical experiments.
|