.png){kind=icon}
Return
Adaptive proximal regularization for image smoothing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.112176.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Image filters based on global optimization have garnered significant attention recently. However, existing optimization models struggle to effectively accommodate the sparsity requirements on the regularization terms. In this paper, we propose to build the regularization term upon the proximal p-norm, which is shown to be able to achieve sparser results. Furthermore, we propose a novel weighting scheme, which adapts to not only the balancing parameter, but also the bandwidth and the sparsity parameters in the proximal p-norm. By adopting different weights, our model provides a general framework for both edge- and structure-preserving image smoothing. Furthermore, in order to overcome the blocky artifacts in two-directional gradient regularization (4-neighborhood), we propose to consider the pixel differences in two more directions (8-neighborhood). Finally, to solve the proposed optimization model with the adaptive non-convex proximal regularization, we propose an efficient solution based on the alternating direction method of multipliers, generalized shrinkage, and the Fourier domain optimization. We have conducted experiments on a variety of applications, including image smoothing, HDR tone mapping, compression artifact removal, texture removal, and depth map upsampling. Both quantitative and qualitative results indicate the superiority of the proposed method. Moreover, our method is highly efficient, it is able to process 720P color images at interactive rates on a modern GPU.
Keywords:
Image smoothing
proximal p-norm
adaptive weighting
non-convex regularization
alternating direction method of multipliers
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
No organization information available
Cited Papers
No cited papers available