Modélisations de textures par champ gaussien à orientation locale prescrite

Kévin Polisano, Marianne Clausel, Valérie Perrier, Laurent CondatarXiv 2015 · DOI · arXiv

Abstract

This paper presents two new models of oriented texture, based on a new class of Gaussian fields, called locally anisotropic fractional Brownian fields (CBFLA), with prescribed local orientation at any point. These fields are a local version of a specific class of anisotropic self-similar Gaussian fields with stationary increments. The simulation of such textures is obtained using a new algorithm mixing the tangent field formulation with the Cholesky method or the turning band method, this latter method having proved its efficiency for generating stationary anisotropic textures. Numerical experiments show the ability of the method for synthesis of textures with prescribed local orientation.

Introduction

Modeling realistic stochastic textures—such as clouds, terrain, or wood grain—is a fundamental challenge in image processing. While Gaussian fields like the Fractional Brownian Field (CBF) are standard tools for creating fractal-like textures, they are typically isotropic, meaning they lack a directional "grain." Existing anisotropic models allow for a global orientation, but they struggle to represent textures where the orientation varies spatially, such as a swirling liquid or a curved wood knot.

This paper introduces Locally Anisotropic Fractional Brownian Fields (CBFLA), a mathematical framework that allows for precise, pixel-level control over texture orientation. By pairing this model with an efficient simulation algorithm, the authors enable the generation of complex textures that follow arbitrary vector fields in seconds.

Methodology: Locally Anisotropic Fields

The core innovation is the extension of Anisotropic Fractional Brownian Fields (CBFA) into a local version. In a CBFA, the texture's orientation is determined by its spectral density, which restricts frequencies to a specific cone in the Fourier domain. The authors evolve this by making the orientation parameter, $\alpha_0$, a function of the spatial position $x$.

To maintain mathematical consistency while allowing spatial variation, the authors utilize the concept of tangent fields. This approach assumes that in a small neighborhood around any point $x_0$, the field behaves like a stationary anisotropic field with a fixed orientation.

Figure 4: (a) Prescribed vector field, (b) local zoom, and (c) diagram showing the spectral density cone used to define local orientation.

Simulation Algorithms

Two primary methods for synthesis are explored:

  1. Cholesky Method: An exact decomposition of the covariance matrix. While highly accurate, its $O(r^6)$ complexity makes it computationally prohibitive for large images.
  2. Turning Band Method: An approximation that represents the 2D field as a sum of 1D fractional Brownian motions projected along various lines. This reduces complexity to $O(r^2 \log n)$, enabling near-instantaneous synthesis of high-resolution textures.

Key Results and Insights

The proposed CBFLA model successfully captures local anisotropy, with synthesized textures "flowing" accurately along prescribed vector fields. The Hurst index ($H$) serves as a critical parameter for texture appearance: a low $H$ (e.g., 0.2) results in rough, jagged textures, while a high $H$ (e.g., 0.7) produces smooth, silk-like surfaces.

Figure 6: Synthesis of a texture following a complex, oscillating vector field, demonstrating the model's ability to handle non-linear local orientations.

The Turning Band method offers a significant speed advantage, generating $256 \times 256$ pixel images in seconds compared to the hours required by the Cholesky method. However, a trade-off exists: the fast algorithm can introduce subtle "streak" artifacts in very smooth textures (high $H$). For rougher textures, these artifacts are negligible.

Conclusion

The CBFLA framework provides a robust method for synthesizing textures with prescribed local orientations. Future work aims to extend this model by allowing the Hurst index itself to vary spatially. Such "Multifractional" fields would enable the simulation of even more complex natural phenomena, where both the regularity and the orientation of the texture change across the image.

Summary updated

Abstract

This paper presents two new models of oriented texture, based on a new class of Gaussian fields, called locally anisotropic fractional Brownian fields, with prescribed local orientation at any point. These fields are a local version of a specific class of anisotropic self-similar Gaussian fields with stationary increments. The simulation of such textures is obtained using a new algorithm mixing the tangent field formulation with the Cholesky method or the turning band method, this latter method having proved its efficiency for generating stationary anisotropic textures. Numerical experiments show the ability of the method for synthesis of textures with prescribed local orientation.
BibTeX
@misc{polisano2015modelisations,
  author = {Kévin Polisano and Marianne Clausel and Valérie Perrier and Laurent Condat},
  title = {{Modélisations de textures par champ gaussien à orientation locale prescrite}},
  year = {2015},
  eprint = {1503.06716},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1503.06716},
}
Figure