Abstract
This paper introduces Sector Diffusion (SD), a novel non-local regularizer designed for multi-frame super-resolution (SR) in the presence of significant noise. By utilizing one-sided directional derivatives, SD preserves sharp edges better than traditional central-difference methods like Edge-Enhancing Diffusion (EED). Furthermore, the study re-evaluates the standard imaging pipeline, discovering that the optimal order of warping, blurring, and downsampling operations shifts when data is corrupted by noise, with the Warping $\rightarrow$ Blurring $\rightarrow$ Downsampling (M1) model proving most robust.
Introduction
Super-resolution (SR) aims to reconstruct high-resolution (HR) images from multiple low-resolution (LR) observations. This inverse problem is inherently ill-posed and becomes increasingly difficult when inputs are corrupted by "clipped" noise, common in biomedical imaging. Traditional methods often rely on simplistic smoothness priors that inadvertently blur critical details. This research introduces Sector Diffusion to address these challenges and investigates how the sequence of operations in the SR observational model affects reconstruction quality in noisy environments.
Methods
The authors propose two primary advancements to improve SR performance under noise:
- Sector Diffusion (SD): Unlike standard diffusion models that use central differences (which can "smear" information across an edge), SD employs one-sided directional derivatives. By dividing a circular neighborhood around a pixel into several "sectors" and performing smoothing only within those sectors, SD avoids crossing edges, thereby maintaining sharp transitions.
- Operator Permutation Analysis: The SR pipeline typically involves three operations: Warping ($W$), Blurring ($B$), and Downsampling ($D$). The authors analyzed six permutations (M1 through M6) to determine which sequence is most robust to additive white Gaussian noise (AWGN) that has been clipped to the standard [0, 255] dynamic range.
Results
Experimental results demonstrate that SD consistently outperforms EED in both denoising and SR tasks.
As shown in Figure 3, SD achieves lower Mean Squared Error (MSE) across all tested images and noise intensities. For instance, on the "Lena" image at $\sigma=40$ (L40), SD reduced MSE to 92.99 from EED's 98.67. This superiority holds even when the input data is heavily corrupted.
Regarding the observational model, the study found that while M2 (Blurring $\rightarrow$ Warping $\rightarrow$ Downsampling) is often preferred in noise-free settings, M1 (Warping $\rightarrow$ Blurring $\rightarrow$ Downsampling) is superior for noisy data.
Figure 5 confirms that Model 1 (M1) yields the lowest MSE in noisy scenarios, whereas the computationally efficient "fast" model (M2.1) suffers significant quality degradation. The authors conjecture that in M1, the errors introduced by swapping non-commuting operators are less damaging than the warping errors exacerbated by noise in other models.
Conclusion
Sector Diffusion provides a robust framework for SR in noisy conditions by leveraging one-sided derivatives to preserve edge integrity. The finding that noise alters the optimal sequence of imaging operators challenges established conventions and highlights the importance of model selection in real-world applications. Future work will focus on reducing the computational overhead of the non-local SD approach and improving its integration with motion estimation techniques.
Summary updated