Source-linked AI summary
On Tensor-Based PDEs and their Corresponding Variational Formulations with Application to Color Image Denoising
Freddie Åström, George Baravdish, Michael Felsberg
TL;DR
The paper asks when a tensor-based diffusion PDE can be represented as the Euler–Lagrange equation of an energy with data and smoothness terms. It derives necessary conditions and applies the resulting formulation to color image denoising, obtaining results comparable to state-of-the-art techniques.
Problem
The paper investigates when a PDE can be the Euler–Lagrange equation of an energy functional with data and smoothness terms, including the challenges of extending diffusion methods to color images.
Method
The paper establishes a tensor-based variational formulation, derives necessary PDE-to-functional conditions, and applies the resulting equation to color image denoising.
Results
The color denoising results are comparable to state-of-the-art techniques, with diffusion preserving structure in high-frequency regions and BM3D favored for approximately homogeneous surfaces.
Takeaways & Limitations
The framework provides a tensor-based variational route for image diffusion and supports color denoising across images with different structural characteristics.
Takeaways & Limitations
The formulation includes choices and approximations: the tensor incorporates a weighting kernel, and the color-noise estimate is biased because the image itself contains noise.
Abstract
from arXiv · showhide
The case when a partial differential equation (PDE) can be considered as an Euler-Lagrange (E-L) equation of an energy functional, consisting of a data term and a smoothness term is investigated. We show the necessary conditions for a PDE to be the E-L equation for a corresponding functional. This energy functional is applied to a color image denoising problem and it is shown that the method compares favorably to current state-of-the-art color image denoising techniques.
1 Introduction
The paper develops conditions linking tensor-based diffusion PDEs to variational energy functionals and applies the resulting Euler–Lagrange equation to color image denoising. It addresses limitations of existing diffusion and variational formulations while comparing favorably with state-of-the-art denoising methods.
- Earlier tensor-based diffusion uses local image orientation, while alternative color-space formulations include manifolds, CMY, HSV, Lab, Luv, luminance–chromaticity, and decorrelation approaches.
- The proposed energy combines an Lp data term, with 1 < p < ∞, and a tensor-based smoothness term.
- Color image denoising remains challenging because the mechanisms governing color perception are not fully known, despite the widespread use of RGB images.RGB components are correlated because of physical properties such as camera color-sensitive filters.
- Prior variational work relates nonlinear diffusion and wavelet shrinkage to variational approaches, but reported tensor-based formulations include extra convolution or static tensors that make the update scheme linear.
- The paper establishes a new tensor-based variational formulation for image diffusion and derives necessary conditions for a tensor-based PDE to have a corresponding energy functional.These results are presented as Theorems 1 and 2.
- The derived Euler–Lagrange equation is applied to color denoising and compared with anisotropic diffusion, trace-based diffusion, and BM3D.
2 Image diffusion and variational formulation
This section reviews scalar and tensor-based diffusion, then derives a tensor-based energy functional whose Euler–Lagrange equation corresponds to a diffusion PDE under explicit conditions. The construction uses image-gradient orientation and supports denoising with data fidelity, boundary conditions, and tensor-dependent smoothness.
- 2 Image diffusion and variational formulation: Gaussian filtering reduces noise variance but ignores edges and lines, motivating Perona–Malik edgestopping and later tensor-based diffusion.
- 2 Image diffusion and variational formulation: Perona–Malik preserves structure but preserves noise at edges because its formulation depends only on the absolute image gradient.
- 2 Image diffusion and variational formulation: Weickert’s tensor replaces scalar diffusivity, reducing noise along edges and lines while preserving image structure through local orientation.
- 2.1 Scalar diffusion: In scalar variational diffusion, a strict convex Φ and scalar λ define a nonlinear image-denoising minimization problem whose minimizer satisfies an Euler–Lagrange equation.
- 2.1 Scalar diffusion: The scalar functional is associated with a PDE initial-boundary value problem using zero normal derivative at the image boundary and initial image u0.
- 2.1 Scalar diffusion: A corresponding scalar functional can be derived by equating an unspecified diffusivity g with the known Euler–Lagrange equation, including examples such as g(s)=s^(p−2) and g(s)=e^(−s).
- 2.2 Deriving a new tensor-based functional: The tensor-based derivation equates the diffusion PDE and Euler–Lagrange equation, allowing a divergence-free vector field C and setting C=0 in this work.
- 2.2 Deriving a new tensor-based functional: The resulting tensor functional generalizes the construction using T(ux,uy), with its Euler–Lagrange equation containing the data term (u−u0)^(p−1), tensor divergence, and Neumann boundary condition.
3 Necessary conditions for the existence of a variational formulation
The paper derives necessary conditions under which a tensor-based diffusion PDE is the Euler–Lagrange equation of a data-plus-smoothness functional. It then constructs tensor formulations, including weighted cases, and establishes their variational correspondence.
- 3 Necessary conditions for the existence of a variational formulation: The section formulates necessary conditions for an energy functional to generate a given tensor-based diffusion PDE as its Euler–Lagrange equation.The framework considers PDEs of the stated tensor-based form and the functional J(u).
- 3 Necessary conditions for the existence of a variational formulation: For a tensor S=S(u_x,u_y) whose entities satisfy the necessary condition, the tensor-based PDE is the Euler–Lagrange equation of J(u).The PDE includes the data term (u−u0)^(p−1), diffusion term −div(S∇u), domain equation, and Neumann boundary condition.
- 3 Necessary conditions for the existence of a variational formulation: The derivation compares expressions for the tensor components and deduces a necessary condition whose right-hand side determines g, after which f and h are obtained.A change of variables α=u_x and β=u_x/u_y reduces the equation for g to αg_α+2g=R.
- 3 Necessary conditions for the existence of a variational formulation: The framework includes the diffusion PDE (u−u0)^(p−1)−div(S∇u)=0 in Ω with ∂_n u=0 on ∂Ω, for 1<p<∞.This places the Lp data term within the variational PDE setting.
- 3 Necessary conditions for the existence of a variational formulation: The proof constructs g by solving the transformed equation, then recovers f and h and verifies that the resulting tensor satisfies the necessary condition.The proof chooses auxiliary functions θ, ρ, and ζ to be zero for simplicity and substitutes the resulting tensor into the functional.
- 3.1 Weighted tensor-based variational formulation: For a weighted tensor S=w*(∇u∇u^T), Proposition 1 states that the associated PDE is the Euler–Lagrange equation of the functional when an additional necessary condition holds.The entities of the auxiliary tensor T can be obtained by solving a system of differential equations.
- 3.1 Weighted tensor-based variational formulation: Including the smooth kernel w in the tensor formulation permits u_x and u_y to be included inside the convolution when sampled at different grid positions.This is the stated rationale for the weighted formulation.
- 3.1 Weighted tensor-based variational formulation: Theorem 1 covers rank(S)=1 structures such as lines, whereas Proposition 1 covers the general rank(S)=2 case before applying the derived equation to color denoising.The section concludes with a framework specifying when a PDE corresponds to a tensor-based image-diffusion energy functional.
4 Application to color image denoising
The denoising application combines a decorrelated color representation with tensor-based diffusion and evaluates the resulting method against several color-image denoising techniques.
- Color representation: The method uses a decorrelation transform whose primary component represents average gray value I and whose other components are color-opponent channels.
- Tensor diffusion: The diffusion tensor scales the eigenvalues of the structure tensor while preserving eigenvectors that encode image-structure orientation.The tensor is constructed from a smoothed outer product of the image gradient and uses a negative exponential scaling function.
- Experimental setup: Noise estimation averages the estimated noise of the three RGB components, but the estimate is biased because the images themselves contain noise.The authors report that this approach generally works reasonably well for noise levels σ < 70.
- Channel-specific parameters: Scaling the diffusion parameter to kI = 10^-1k and kc = 10k improves results compared with using k without channel-specific scaling.The gray-value channel is assigned a smaller parameter, while color-opponent filtering uses a less structure-preserving parameter to reduce color artifacts.
- Evaluation: Experiments compare the proposed technique with trace-based diffusion, RGB-space diffusion, and color BM3D using averaged RGB SSIM and PSNR measurements.Gaussian noise with standard deviations {5, 10, 20, 50, 70} is added before filtering.
- Results: The established E-L equation performs well overall, while BM3D is favored on approximately homogeneous surfaces and diffusion performs better in high-frequency regions.The figure is reported to show structure preservation by diffusion methods on the lizard, snake, and owl images, with perceptually good BM3D results for the mushroom image.
5 Conclusion
The paper establishes necessary conditions for tensor-based image-diffusion PDEs to correspond to energy-function Euler–Lagrange equations and applies the resulting equation to color denoising.
- The work gives necessary conditions for a tensor-based image-diffusion PDE to be the Euler–Lagrange equation of an energy functional.
- The derived Euler–Lagrange equation is applied to color-image denoising with results comparable to state-of-the-art techniques.The conclusion states that generalization of Proposition 1 and its complete proof remain subject to further study.