Telea vs. Navier-Stokes Inpainting Algorithms
When a digital watermark has solid, opaque regions ($alpha o 1.0$), reverse alpha subtraction faces a mathematical singularity. In these opaque zones, computer vision engines must turn to edge-directed inpainting. We analyze the two premier algorithms: Alexandru Telea’s Fast Marching Method and Navier-Stokes Partial Differential Equations.
1. The Goal of Non-Generative Inpainting
Digital inpainting aims to reconstruct missing or damaged regions ($Omega$) inside an image domain such that observers cannot detect the intervention. Unlike generative AI models that fabricate new objects out of probabilistic noise, structural inpainting propagates existing geometric edges, textures, and gradient fields continuously inward from the boundary ($partialOmega$).
2. Alexandru Telea’s Fast Marching Method (FMM)
Published in 2004, Telea’s algorithm models inpainting as the propagation of an interface using the Eikonal equation $| abla T| = 1$, where $T(x, y)$ represents the geodesic distance of point $(x, y)$ to the mask boundary $partialOmega$.
Pixels are synthesized in order of increasing distance from the boundary using the Fast Marching Method. To calculate the color $I(p)$ of an interior pixel $p$ from known neighborhood pixels $q in B_epsilon(p)$:
I(p) = Σ_(q ∈ B_ε(p)) w(p, q) × [ I(q) + ∇I(q) × (p - q) ] / Σ_(q ∈ B_ε(p)) w(p, q)
The weighting kernel $w(p, q)$ combines directional alignment, geometric distance, and level-set propagation:
w(p, q) = dir(p, q) × dst(p, q) × lev(p, q)
Telea excels at speed ($O(N log N)$) and delivers smooth, clean reconstructions across thin scratches and typographic contours.
3. Navier-Stokes Fluid Dynamics Inpainting (Bertalmio et al.)
Developed by Marcelo Bertalmio, Andrea Bertozzi, and Guillermo Sapiro (2001), this approach connects image inpainting with 2D fluid dynamics. The intensity of an image is modeled as a stream function in an incompressible, inviscid 2D fluid flow, where isophote lines (lines of equal intensity) correspond to streamlines.
The transport of image information is governed by the vorticity transport equation:
∂ω / ∂t + v × ∇ω = ν Δω
Where $omega = Delta I$ represents the smoothness (Laplacian/vorticity) of the image, $v = abla^perp I = (-partial I / partial y, partial I / partial x)$ represents the perpendicular isophote direction vector, and $ u Deltaomega$ represents an anisotropic diffusion term that smooths noise across streamlines.
4. Comparative Tradeoffs for Watermark Erasing
| Feature | Telea Fast Marching (FMM) | Navier-Stokes (PDE) |
|---|---|---|
| Primary Mechanism | Eikonal Distance-Weighted Marching | Isophote Streamline Fluid Continuation |
| Edge Preservation | Good along moderate contours | Exceptional across strong continuous edges |
| Computational Complexity | Very Fast ($O(N log N)$) | Moderate (Iterative Jacobi/SOR steps) |
| Best Application | Thin text, fine logos, small star arms | Broad solid badges, complex photographic horizons |
Test Both Algorithms in Real Time
AURA ERASE lets you seamlessly switch between Reverse Alpha Deblending, Telea Inpainting, and Navier-Stokes.
Configure Inpainting in Studio →