Deep K-SVD for image denoising

This thesis aims to implement a baseline version of tghe deep K-SVD algorithm for the denoising of images. It then explore the possible enhancement on the performance using a training function which improve the common Mean Squared Error, combining it with a function which takes into account structure similarity between the output image and the original one. The mixed function need then to be regularized using a criterion based on the Normalized Cumulative Periodogram applied to the residual, and comparing it with the added white noise.