TL;DR
This paper introduces the concept of "algorithm-induced priors" for image restoration, where the prior is implicitly defined by the denoising step within an ADMM algorithm, rather than being explicitly defined beforehand. The authors derive an explicit mathematical form for these priors when using symmetric smoothing filters, showing they offer improved performance over traditional graph Laplacian priors and can lead to efficient closed-form solutions.
What Problem Does This Paper Address?
Image restoration tasks, like denoising or inpainting, often rely on "priors" – assumptions about the expected properties of a clean image (e.g., smoothness). These priors are typically defined mathematically before an optimization algorithm like ADMM is run. However, a popular and effective approach is the "plug-and-play" ADMM, where a pre-trained denoising algorithm is used within the ADMM framework. This raises a fundamental question: what is the actual mathematical prior that this plug-and-play approach is implicitly enforcing? Understanding this "algorithm-induced prior" is crucial for analyzing its performance and limitations, but it has been difficult because it's not explicitly stated. This paper aims to bridge this gap by deriving the explicit form of these algorithm-induced priors.
Methods
The core of the paper lies in analyzing the ADMM algorithm, specifically when its denoising module is replaced by an off-the-shelf denoising algorithm. The authors first reformulate the standard ADMM iterations to highlight its modular structure, separating the "inverse" and "denoising" steps.
They then focus on a specific class of denoising algorithms: symmetric smoothing filters. For these filters, they derive an explicit mathematical expression for the algorithm-induced prior. This new prior is shown to be related to the conventional graph Laplacian but incorporates an additional term derived from the pseudo-inverse of the smoothing filter matrix.
Furthermore, the paper presents a method to derive a closed-form solution for the image restoration problem when using these algorithm-induced priors. This is achieved by leveraging the eigen-decomposition of the smoothing filter matrix, which allows for a more computationally efficient solution compared to directly inverting large matrices.
Datasets
Not discussed in the paper. The experiments use synthetic data for image inpainting and deblurring, with the "ground truth" being the original clean image.
Key Insights and Ideas
- The "Plug-and-Play" has a Prior: The key insight is that using a denoising algorithm within ADMM isn't arbitrary; it implicitly defines a specific image prior. This prior is not necessarily the one the denoising algorithm was originally designed for, but rather what it enforces when combined with the ADMM framework.
- Symmetric Smoothing Filters Yield a Novel Prior: By analyzing symmetric smoothing filters, the authors derive a new prior that is mathematically distinct from the standard graph Laplacian. This new prior is shown to be more effective in practice.
- Algorithm-Induced Prior as a Bandpass Filter: The derived prior, using the pseudo-inverse of the smoothing filter, acts as a "bandpass filter" on image frequencies. This is in contrast to the conventional graph Laplacian, which acts more like a "highpass filter." This difference explains why the new prior can be more effective by penalizing a more focused set of frequencies.
- Efficient Closed-Form Solutions: The paper demonstrates that for certain symmetric smoothing filters, the resulting optimization problem can be solved efficiently in closed-form, bypassing the iterative nature of ADMM. This is achieved by exploiting the matrix structure and eigen-decomposition.
Key Results
- Improved Performance: Experiments on image inpainting show that the proposed algorithm-induced prior consistently outperforms the conventional graph Laplacian prior, especially in scenarios with significant missing pixels. For example, at 80% missing pixels, the proposed prior achieved a PSNR of 21.85dB compared to 21.56dB for the conventional Laplacian.
- Visual Improvement: Figure 1 provides a visual comparison, demonstrating that the algorithm-induced prior leads to visually superior results in image inpainting compared to the conventional graph Laplacian.
- Closed-Form Solution Derivation: The paper successfully derives an efficient closed-form solution for image restoration problems using these algorithm-induced priors, which is computationally advantageous.
Limitations and Future Work
- Symmetric Smoothing Filters Focus: The explicit derivation of the prior is currently limited to the class of symmetric smoothing filters. The paper suggests that extending this analysis to other types of denoising algorithms would be valuable.
- Parameter Tuning: While the paper shows that optimal parameters differ for different priors, it doesn't provide a general method for automatically selecting the best parameters for the algorithm-induced prior.
- Understanding the Interplay: The authors emphasize the need for further research into the interplay between the objective function (forward model) and the denoising procedure to gain a deeper understanding of these algorithms.
- Generalization to Other Tasks: While demonstrated for inpainting, exploring the effectiveness of algorithm-induced priors for other image restoration tasks like deblurring and super-resolution is a natural next step.
Summary updated