Abstract: This paper presents a new framework for oriented texture modeling. We introduce a new class of Gaussian fields, called Locally Anisotropic Fractional Brownian Fields (LAFBF), 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 and a 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
In image processing and computer graphics, modeling natural textures remains a significant challenge. Many natural surfaces—such as clouds, rocky terrain, or waves—exhibit self-similarity, meaning they maintain a consistent appearance across different scales. The standard mathematical tool for modeling such phenomena is the Fractional Brownian Field (FBF).
However, standard FBFs are isotropic, meaning their statistical properties are identical in every direction. Real-world textures often possess a distinct "grain" or directionality, known as anisotropy. While existing models allow for a single global orientation across an entire image, they struggle to represent textures where the orientation curves or changes locally, such as wood grain, combed fur, or wind-blown sand. This paper addresses this gap by introducing a model where anisotropy can be prescribed as a varying vector field across the image.
Methodology
Locally Anisotropic Fractional Brownian Fields (LAFBF)
The authors define a new class of Gaussian fields where the spectral density (the frequency content) is modified by a local orientation parameter $\alpha_0(x)$. This formulation effectively creates a local version of an Anisotropic Fractional Brownian Field (AFBF). To ensure mathematical tractability, the model utilizes the concept of tangent fields. This approach assumes that at any specific point $x$, the complex, spatially varying field behaves locally like a simpler, stationary anisotropic field.
Efficient Simulation: The Turning Band Method
Simulating Gaussian fields traditionally involves massive computational costs due to the need to manipulate large covariance matrices. To overcome this, the authors employ the turning band method. This technique decomposes a 2D field into a sum of independent 1D Fractional Brownian Motions projected along various angular "bands."
The resulting algorithm is highly efficient, with a computational complexity of $O(r^2 \log n)$, where $r$ is the image resolution and $n$ is the number of angular bands. This allows for the synthesis of high-resolution textures in just a few seconds on standard hardware.
Key Insights and Results
The Orthogonality Effect
A critical observation in the synthesis process is the relationship between the prescribed vector field and the resulting texture. Due to the properties of the Fourier transform, the synthesized patterns appear visually orthogonal to the prescribed vector field. Consequently, to make a texture "flow" along a specific path, the user must prescribe a vector field that is perpendicular to that path.
Visual Fidelity and Control
The model provides explicit control over two primary features:
- Hurst Index ($H$): Controls the "roughness" or "smoothness" of the texture.
- Local Orientation ($\alpha_0$): Controls the directionality of the grain.
As demonstrated in Figure 6, the synthesized textures successfully follow complex, non-linear orientations while maintaining the characteristic fractal roughness of Brownian motion. This capability allows for the creation of organic, flowing patterns that were previously difficult to generate using purely stochastic models.
Limitations and Future Work
While the current implementation is robust, the authors identify a few areas for improvement:
- Numerical Artifacts: At higher values of the Hurst index ($H$), discrete implementations can produce artifacts. Future work will focus on improving the regularization of the indicator functions used in the turning band method.
- Spatially Varying Roughness: Currently, the Hurst index $H$ is constant across the image. A natural extension is to allow both $H$ and the orientation to vary spatially, enabling textures that are rougher in some regions and smoother in others.
- Image Decomposition: The authors suggest this model could be applied to "cartoon+texture" image decomposition, helping to separate structural image components from underlying oriented textures.
Summary updated