Variable-order fractional 1-Laplacian diffusion equations for multiplicative noise removal

Yuhang Li, Zhichang Guo, Jingfeng Shao, Yao Li, Boying WuFractional 2024 · DOI

TL;DR

This paper introduces a novel variable-order fractional 1-Laplacian diffusion equation designed to tackle multiplicative noise in images. By allowing the order of the fractional Laplacian to vary spatially, the model achieves enhanced adaptability, leading to more effective noise removal and better preservation of image textures compared to existing methods.

What Problem Does This Paper Address?

Image denoising is a critical task in image processing, and multiplicative noise, where the observed image is a product of the original image and noise, presents a particular challenge. Traditional diffusion equations, like the heat equation or p-Laplacian equations, are effective for additive noise but struggle with multiplicative noise. While some models have been proposed, they often suffer from either over-smoothing, leaving residual noise, or failing to preserve fine image details like textures and repetitive structures. Existing nonlocal methods, while better at preserving textures, can still lead to over-smoothing in low-contrast regions and leave noise around edges. This paper aims to address these limitations by proposing a new diffusion model that can adapt its denoising behavior based on local image characteristics.

Methods

The core of this work is the introduction of a variable-order fractional 1-Laplacian diffusion equation. Unlike standard fractional diffusion models where the order is constant, this model allows the order $s(x,y)$ to vary depending on the spatial location $(x,y)$. This spatial variation is key to the model's adaptability.

The authors leverage Nonlinear Semigroup Theory to prove the existence and uniqueness of weak solutions for their proposed equation. This involves defining and analyzing variable-order fractional Sobolev spaces ($W^{s(\cdot,\cdot),p}(\Omega)$) and their properties, including density and embedding theorems.

The model is formulated as a reaction-diffusion equation with a specific diffusion term involving the variable-order fractional 1-Laplacian operator. The weight function $k(x,y)$ is also designed to be adaptive, further enhancing the model's performance. The authors also propose specific choices for the variable order function $s(x,y)$ and the weight function $k(x,y)$ that are inspired by texture analysis (using Gabor filters) and local image properties.

Datasets

The paper evaluates the proposed model on several standard test images:

  • Texture1
  • Texture2
  • Hybrid
  • Satellitel
  • Satellite2

These images are corrupted with multiplicative Gamma noise with varying levels of noise intensity ($L = 1, 4, 10$). No new datasets are introduced.

Key Insights and Ideas

  • Adaptive Denoising through Variable Order: The most significant insight is that allowing the order of the fractional Laplacian to vary spatially ($s(x,y)$) enables the diffusion process to adapt to local image content. This means the model can diffuse more strongly in smooth regions to remove noise while preserving edges and textures by diffusing less.
  • Texture Preservation: By carefully designing $s(x,y)$ to be close to 1 in texture-poor regions and closer to a non-extreme order in texture-rich regions, the model effectively preserves fine details that are often lost by other methods.
  • Absence of Source Term: Unlike some previous models for multiplicative noise removal, this approach avoids a source term, which can sometimes hinder effective denoising. The adaptive diffusion coefficient is relied upon for noise removal.
  • Theoretical Foundation: The paper provides a rigorous mathematical framework by proving the existence and uniqueness of weak solutions using advanced tools like Nonlinear Semigroup Theory and variable-order fractional Sobolev spaces.

Key Results

The paper demonstrates the effectiveness of the proposed model through numerical experiments and quantitative metrics:

  • Quantitative Comparison: The proposed model (labeled "Ours") consistently achieves higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) values compared to existing methods like SO, AA, and F1P-AA across various test images and noise levels. For example, in Table 1, for the "Texture1" image with $L=4$, the proposed model achieves a PSNR of 15.68, outperforming SO (15.35), AA (15.52), and F1P-AA (15.51).
  • Qualitative Improvements: Visual comparisons (Figures 2-6) show that the proposed model excels at removing multiplicative noise while preserving image textures and edges. It avoids the over-smoothing seen in some other methods and better retains fine details. The model achieves a good balance between noise removal and texture preservation.

Limitations and Future Work

  • Parameter Selection: The paper acknowledges that the performance of the model relies on the appropriate selection of parameters for $k(x,y)$ and $s(x,y)$, including constants related to Gabor filters and the variable order function. While they suggest methods for parameter selection, finding optimal parameters for diverse images might still be challenging.
  • Computational Cost: While not explicitly detailed as a limitation, fractional diffusion models, especially with variable orders and nonlocal operators, can be computationally intensive. The paper focuses on theoretical aspects and numerical experiments, but efficiency for real-time applications might be a future consideration.
  • Future Work: The authors suggest that further research could explore different choices for the variable order function $s(x,y)$ and the weight function $k(x,y)$ to further enhance the model's adaptability and performance. Investigating the application of this model to other types of noise or image processing tasks is also a potential avenue.

Summary updated

BibTeX
@inproceedings{li2024variable,
  author = {Yuhang Li and Zhichang Guo and Jingfeng Shao and Yao Li and Boying Wu},
  title = {{Variable-order fractional 1-Laplacian diffusion equations for multiplicative noise removal}},
  booktitle = {Fractional},
  year = {2024},
  doi = {10.1007/s13540-024-00345-6},
}
Figure