Abstract:Anisotropic diffusion processes with a diffusion tensor are important in image analysis, physics, and engineering. However, their numerical approximation has a strong impact on dissipative artefacts and deviations from rotation invariance. In this work, we study a large family of finite difference discretisations on a 3 x 3 stencil. We derive it by splitting 2-D anisotropic diffusion into four 1-D diffusions. The resulting stencil class involves one free parameter and covers a wide range of existing discretisations. It comprises the full stencil family of Weickert et al. (2013) and shows that their two parameters contain redundancy. Furthermore, we establish a bound on the spectral norm of the matrix corresponding to the stencil. This gives time step size limits that guarantee stability of an explicit scheme in the Euclidean norm. Our directional splitting also allows a very natural translation of the explicit scheme into ResNet blocks. Employing neural network libraries enables simple and highly efficient parallel implementations on GPUs.

What problem does this paper attempt to address?

The main focus of this paper is on numerical approximation methods for anisotropic diffusion in image analysis. Anisotropic diffusion has wide applications in physics, engineering, and image analysis, especially in tasks such as denoising, enhancement, scale-space analysis, and interpolation. The paper proposes a new 3x3 stencil by decomposing 2D anisotropic diffusion into four 1D diffusion processes to simplify the numerical approximation, reduce distortion effects, and improve rotational invariance. Specifically, the contributions of the paper are as follows: 1. Investigation of a large family of finite difference discretizations on a 3x3 stencil, which decomposes 2D anisotropic diffusion into four 1D diffusion processes. This family has a free parameter that can optimize quality, covers the previously proposed two-parameter stencil family by Weickert et al., and reveals its parameter redundancy. 2. Provision of limits on the matrix spectral norm, which gives time step constraints to ensure stability of explicit schemes under Euclidean norm. 3. Transformation of the explicit anisotropic diffusion scheme into ResNet blocks, demonstrating the shared ideas between this numerical algorithm and neural network architectures, and implementation of simple and efficient parallelization on GPUs using a neural network library. The paper first introduces the 1D diffusion model and its discretization, and then extends it to 2D anisotropic diffusion, proposing the δ-stencil as a direction-splitting approach. Stability analysis is then conducted to determine the time step limits of the explicit scheme. Finally, the paper demonstrates how this method can be transformed into a ResNet architecture for fast parallel execution on GPUs.

Anisotropic Diffusion Stencils: From Simple Derivations over Stability Estimates to ResNet Implementations

Analysis and Design of Anisotropic Diffusion for Image Processing

Behavioral Analysis of Anisotropic Diffusion in Image Processing .

STENCIL-NET: Data-driven solution-adaptive discretization of partial differential equations

An Arbitrary N-dimensional Stencil Transformer in Cilk + + ∗

Reverse Transition Kernel: A Flexible Framework to Accelerate Diffusion Inference

Translating Numerical Concepts for PDEs into Neural Architectures

STENCIL-NET for equation-free forecasting from data

Connections Between Numerical Algorithms for PDEs and Neural Networks

Improved Stencil Selection for Meshless Finite Difference Methods in 3D

AdjointDEIS: Efficient Gradients for Diffusion Models

StableDreamer: Taming Noisy Score Distillation Sampling for Text-to-3D

DiffusionNet: Discretization Agnostic Learning on Surfaces

Isotropic optimizations of finite difference discretization

A Tunable Despeckling Neural Network Stabilized via Diffusion Equation

Diffusion bridges for stochastic Hamiltonian systems and shape evolutions

simple diffusion: End-to-end diffusion for high resolution images

Anisotropic Diffusion on Curved Surfaces

DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models

StEik: Stabilizing the Optimization of Neural Signed Distance Functions and Finer Shape Representation

Temporal Vectorization for Stencils