Analysis of sparsity- and nonlocality-reinforced convolutional neural networks for image denoising

Abstract

Different neural networks are built using the same elements, but they vary a lot by architecture and show different denoising quality results. The master thesis aims to analyze how different network hyperparameters, input transformations (sparsity), and nonlocal filters (nonlocality) impact the performance in image denoising tasks. The thesis work provides rich experimentally based research and analysis in the image denoising field. A number of different denoising methods have been considered, and state-of-the-art denoising algorithms were examined. As a result of the parameter analysis, the vast majority of the learning-based algorithms and deep networks were improved in terms of denoising capability.