Abstract
In this work, we broadly connect kernel-based filtering (e.g. approaches such as the bilateral filters and nonlocal means, but also many more) with general variational formulations of Bayesian regularized least squares, and the related concept of proximal operators. The latter set of variational/Bayesian/proximal formulations often result in optimization problems that do not have closed-form solutions, and therefore typically require global iterative solutions. Our main contribution here is to establish how one can approximate the solution of the resulting global optimization problems with use of locally adaptive filters with specific kernels. Our results are valid for small regularization strength but the approach is powerful enough to be useful for a wide range of applications because we expose how to derive a "kernelized" solution to these problems that approximates the global solution in one-shot, using only local operations. As another side benefit in the reverse direction, given a local data-adaptive filter constructed with a particular choice of kernel, we enable the interpretation of such filters in the variational/Bayesian/proximal framework.
Introduction
Image processing has long been divided between two powerful paradigms: variational/Bayesian methods and local kernel-based filtering. Variational methods use explicit priors (like Total Variation) to define clean images, requiring computationally expensive global iterative optimization. Conversely, local filters like the Bilateral Filter or Non-Local Means are fast and "one-shot" but often lack a rigorous statistical foundation. This paper establishes a mathematical bridge between these worlds, demonstrating how complex global priors can be approximated by fast, local operations.
Theoretical Framework
The core of the proposed approach lies in the relationship between proximal operators and the graph Laplacian. In a Bayesian context, image restoration often involves finding the Maximum A Posteriori (MAP) estimate, which corresponds to the peak of the posterior distribution.
For small noise levels ($\sigma^2$), the authors demonstrate that the MAP estimate can be approximated by a single "explicit Euler step": $\hat{x} \approx x - \sigma^2 \nabla \phi(x)$, where $\phi$ is the regularization function. By relating the gradient of the loss function ($\nabla \phi$) to the graph Laplacian ($L$), the authors establish a First-Order Relationship: $\nabla \phi(x) \approx \frac{1}{\sigma^2} L(x)x$. Since the Laplacian is defined by filter weights ($L = I - W$), this allows for the direct mapping of a specific global loss function to a local kernel.
Key Insights
- Robustness of Common Kernels: A significant finding is that the standard Gaussian kernel implies a "robust" Welsch loss function, explaining why Gaussian-based filters effectively ignore outliers.
- Local Total Variation: The computationally heavy Total Variation (TV) prior can be implemented as a local filter using a $1/|t|$ kernel.
- One-Shot Efficiency: For small regularization parameters, a single pass of a "kernelized" global prior behaves almost identically to the ideal global solution, enabling high-quality restoration at mobile-friendly speeds.
Results and Analysis
The framework's accuracy was validated using standard image processing benchmarks. In denoising experiments, a first-order local approximation of Huber-Total Variation achieved a PSNR of 37.10 dB, nearly matching the iterative global solution. Similarly, approximating the Bilateral Filter as a MAP problem yielded a high PSNR of 48.48 dB for $\sigma=30$.
As shown in the figure above, the approximation is highly accurate for small $\sigma$. However, as the noise level or regularization strength increases, the approximation error grows, indicating where iterative methods may still be necessary.
Conclusion
This work provides a principled way to derive local filters from global objectives and, conversely, to "unmask" the underlying Bayesian priors of existing heuristic filters. While the "one-shot" approximation is most effective for small noise levels, it offers a powerful tool for real-time image restoration. Future work will focus on refining higher-order approximations and addressing the complex inverse problem of finding valid loss functions for arbitrary kernels.
Summary updated