TL;DR
This paper extends Diffusion-Shock (DS) filtering, a powerful image denoising and inpainting technique, to the space of positions and orientations ($M_2$). This novel approach allows for the preservation of crossing structures in images, a significant limitation of previous DS filtering methods. The authors introduce two new variants, left-invariant and gauge frame RDS filtering, demonstrating their superior performance in denoising and inpainting tasks, particularly for images with complex structures.
What Problem Does This Paper Address?
Image processing techniques like denoising and inpainting often struggle with images containing complex structures, such as crossings or overlapping lines. Traditional methods, including the Regularised Diffusion-Shock (RDS) filtering, operate in the standard Euclidean space ($R^2$) and tend to smooth out or incorrectly connect these intricate features. This limitation hinders their ability to accurately restore or enhance such images. The paper addresses this by proposing an extension of RDS filtering to the space of positions and orientations ($M_2 = R^2 \times S^1$). This lifted space allows for a more nuanced representation of image features, where orientations are explicitly considered alongside spatial locations, thereby disentangling crossing structures and enabling their proper processing.
Methods
The core contribution of this paper is the extension of Regularised Diffusion-Shock (RDS) filtering to the space of positions and orientations ($M_2$). This involves several key methodological advancements:
- Lifting to $M_2$: Images are transformed into the $M_2$ space using an "orientation score transform" ($W_\psi$), which utilizes cake wavelets. This process disentangles crossing structures by representing them with distinct orientations.
- Left-Invariant RDS Filtering: This variant utilizes left-invariant vector fields on $M_2$. The diffusion component is based on the Lie-Cartan Laplacian, which, for unimodular Lie groups like $SE(2)$ (the group acting on $M_2$), coincides with the Laplace-Beltrami operator.
- Gauge Frame RDS Filtering: This approach addresses practical limitations arising from discretizing orientations. It employs "gauge frames" that are adapted to the local image data, mitigating issues caused by the finite number of orientations. The diffusion in this case is driven by a "data-driven Lie-Cartan Laplacian," which is shown to differ from the standard Laplace-Beltrami operator.
- Coherence-Enhancing Shock Filtering: Both variants incorporate a shock filtering component that sharpens and elongates edge-like structures, adapted for the $M_2$ space. This is achieved by using a sigmoidal function applied to a Laplacian perpendicular to the local orientation.
- Regularisation: The paper discusses the use of Gaussian kernels for regularizing derivatives in the guidance terms, ensuring stability.
Datasets
The paper does not introduce new datasets. Instead, it evaluates its methods on existing image datasets, specifically:
- A medical image patch of the retina.
- A cartoon-like image of overlapping spirals.
These images are degraded with additive, correlated noise for denoising experiments.
Key Insights and Ideas
- Disentangling Crossings: Lifting image data to the space of positions and orientations ($M_2$) is crucial for disentangling crossing structures. This allows for processing that respects the distinct paths of intersecting lines, which is impossible in $R^2$.
- Gauge Frames for Practicality: While the theoretical framework of left-invariant vector fields is elegant, practical implementations often require adapting to data-specific orientations. Gauge frames provide a data-driven way to achieve this, leading to improved performance.
- Laplacian Differences: The paper highlights a theoretical distinction between the Laplacians used in the left-invariant and gauge frame settings. While they coincide for unimodular Lie groups in the left-invariant case, the gauge frame setting introduces a "data-driven" Laplacian that differs from the standard Laplace-Beltrami operator. This difference is important for understanding the underlying diffusion process.
- Inpainting Capability: The $M_2$ extension of RDS filtering enables inpainting of crossing structures, a capability that the $R^2$ version lacks. This is a significant advancement for image restoration tasks involving complex geometries.
Key Results
- Denoising Performance: The proposed $M_2$ RDS filtering methods (both left-invariant and gauge frame) outperform existing techniques, including $R^2$ RDS filtering, Total Roto-Translational Variation (TR-TV) flows on $M_2$, BM3D, and NLM, in terms of Peak Signal-to-Noise Ratio (PSNR).
- Robustness: The $M_2$ RDS filtering methods are less sensitive to the stopping time compared to methods like BM3D and NLM.
- Inpainting of Crossings: Figure 6 visually demonstrates that $M_2$ RDS inpainting successfully reconstructs crossing lines, whereas $R^2$ RDS inpainting connects them without preserving the crossing.
- Qualitative Improvements: Visual comparisons (Figures 4 and 5) show that PDE-based methods on $M_2$ preserve fine structures (like small vessels) better than NLM, and the gauge frame RDS filtering performs slightly better than the left-invariant version.
Limitations and Future Work
- Limitations: The paper acknowledges that the gauge frame approach involves a finite number of orientations, which can lead to a "deviation from horizontality" and requires careful adaptation. The theoretical analysis of the gauge frame diffusion's well-posedness and smoothing properties is noted as an area for future investigation.
- Future Work: The authors plan to integrate this crossing-preserving shock filtering into PDE-based deep learning architectures, building on prior work in geometric image processing on $M_2$.