New Feature Generation for Denoising Gradient-Domain Rendering

Abstract

In this thesis, we propose a new technique to utilize a feature-based denoiser in a gradient-domain rendering framework efficiently. The main idea of our method is to make a new feature using image gradients that originate from the gradient-domain rendering and put the generated feature into local regression based adaptive sampling and reconstruction method, which uses several rendering-specific features such as normal. The ultimate goal of this method is that some noise and artifacts can be additionally removed in the reconstructed image by making use of our generated feature with other rendering-specific features into the local regression based method. Moreover, for the purpose of robust estimation of our feature, we introduce a robust feature generation process including a weighted least squares (WLS) and bagging process.We demonstrate that our solution makes a quality improvement by linking the local regression based method into the gradient-domain rendering framework seamlessly.