Abstract
This paper addresses the problems of (1) representing natural shapes such as mountains, trees and clouds, and (2) computing such a description from image data. In order to solve these problems we must be able to relate natural surfaces to their images; this requires a good model of natural surface shapes. Fractal functions are good a choice for modeling natural surfaces because (1) many physical processes produce a fractal surface shape, (2) fractals are widely used as a graphics tool for generating natural-looking shapes, and (3) a survey of natural imagery has shown that the 3-D fractal surface model, transformed by the image formation process, furnishes an accurate description of both textured and shaded image regions. This characterization of image regions has been shown to be stable over transformations of scale and linear transforms of intensity.
Much work has been accomplished that is relevant to computing 3-D information from the image data, and the computation of a 3-D fractal-based representation from actual image data has been demonstrated using an image of a mountain. This example shows the potential of a fractal-based representation for efficiently computing good 3-D representations of natural shapes, including such seemingly-difficult cases as mountains, clumps of leaves and clouds.
Introduction
In the early 1980s, computer vision faced a significant hurdle: the "messiness" of the real world. While traditional computer science attempted to describe the environment using smooth Euclidean primitives—spheres, cylinders, and cubes—these models failed to capture the complexity of natural scenes. Alex Pentland’s 1984 research proposed a paradigm shift, arguing that nature speaks the language of fractals rather than smooth geometry.
Standard primitives become implausibly complex when describing something as common as a crumpled newspaper or a jagged mountain range. Furthermore, there was no rigorous physical model to relate the 3D roughness of a surface to the 2D texture observed in a photograph. Pentland addressed these gaps by introducing a framework using Fractal Brownian functions to model, describe, and segment natural shapes.
The Fractal Model of Natural Surfaces
Pentland leverages Fractal Brownian Motion (fBm), a mathematical framework popularized by Benoit Mandelbrot, to provide a formal measure of "jaggedness" known as the Fractal Dimension ($D$). In this model, a flat plane has a dimension of $D=2$, while a highly rugged, space-filling surface approaches $D=3$. This metric maps effectively to human intuition: a $D$ value of approximately 2.1 corresponds to rolling countryside, while $D \approx 2.5$ represents young, rugged mountains.
The paper’s most significant theoretical contribution is the link between 3D surfaces and 2D image formation. Pentland proves that if a 3D surface is a spatially isotropic fractal, the resulting image intensity surface is also fractal. This allows researchers to estimate 3D roughness directly from 2D pixel data using two primary methods:
- Fourier Power Spectrum: Measuring the rate at which power falls off as frequency increases.
- Second-Order Statistics: Measuring how the variance of point-to-point intensity differences changes with distance.
Experimental Results and Segmentation
The model was evaluated using a variety of natural scenes, including aerial photography of San Francisco Bay, desert landscapes, and the standard Brodatz texture benchmark. The fractal-based segmenter achieved an 84.4% classification accuracy on the Brodatz texture mosaic, significantly outperforming then-standard correlation techniques (65%) and co-occurrence methods (72%).
One of the most robust features of this approach is its scale invariance. Because fractals look statistically similar at different scales, the fractal dimension remains a stable descriptor whether an object is viewed from one mile or ten miles away. In tests on aerial imagery, the segmentation of land versus water remained consistent despite significant variations in scale.
Discussion and Limitations
Pentland’s work suggests a "constrained-chance" representation of the world. Rather than storing the exact $Z$-coordinate of every point on a mountain, we can store the underlying "rules" (the fractal dimension and boundary conditions) and use stochastic processes to fill in the details. This philosophy implies that we do not need to know the position of every leaf in a bush to recognize it; we only need to understand the process that created it.
However, the fractal model is not universal. It fails for regular, man-made structures like brick walls or tiled floors, which require structural rather than stochastic descriptions. Additionally, the basic model assumes isotropy (the same texture in all directions), and computing dimensions near texture boundaries can occasionally lead to edge artifacts. Despite these limitations, the paper demonstrated that 77% of natural textures could be accurately described by the fractal model, providing a powerful tool for the automated interpretation of the natural world.
Summary updated