.png){kind=icon}
.png){kind=icon}
Return
A variable parameter variational model with application to real image denoising
.png){kind=icon}
PRE
AI
DOI：10.1016/j.sigpro.2024.109593.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper establishes AdaTV, a flexible variational framework with learnable adaptive parameters realized by a deep CNN, to effectively tackle the challenge of real image denoising problem. The AdaTV enjoys the benefits of variational model in describing image statistic prior via maximize a posterior (MAP), meanwhile, it takes full advantages of CNN to specify the adaptive parameters of the variational model. Specifically, the variational model consists of a tailored total variation (TV) model with some learnable parameters achieved by a deep CNN. Since in real image denoising it is difficult to obtain high quality paired training data (noisy-clean training pairs), in AdaTV we design a novel unpaired bi-level learning model to learn such a specialized deep CNN. To mitigate the computational complexity, we further relax the bi-level learning model into a Nash game, by solving the game model, we can train the deep CNN in more efficient way. Theoretical analysis and numerical experiments consistently demonstrate that AdaTV exhibits superior interpretability, robustness, generalization, and performance on real image denoising problem.
Keywords:
Real image denoising
Total variation
Unpaired learning
Parameters learning