Modeling enhancements in the DUDE framework for grayscale image denoising

摘要

We present recent theoretical and practical developments aimed at enhancing the performance of the discrete universal denoiser (DUDE) [1] on grayscale images. In particular, a new statistical model for images, formalizing the assumptions underlying the use of prediction, together with a more robust use of pre-filtering and iteration have led to significant improvements in denoising performance for certain types of noise, compared with the state of the art (which includes the first DUDE implementation for this application in [2]).