45  Fractals, Scale Invariance, and Natural Textures

Published

September 17, 2026

Work in Progress

Suggestions of all kinds for this book draft are welcome — whether it’s fixing small errors, raising bigger questions, or offering new perspectives. Please share comments through GitHub Issues. To make feedback easier to address, please point to the section you have in mind — by section number or a short snippet of text.

Return to Spatial Properties of Light Fields | Return to book home

45.1 Overview and Historical Context

In Chapter 5 (Spatial Properties), we saw that natural scenes generally exhibit a scale-invariant power spectrum falling off roughly as \(1/f^\alpha\) (typically with \(\alpha \approx 2\)). Historically, this empirical observation triggered immense excitement across computer vision, computer graphics, and visual psychophysics. If natural scenes lack a characteristic spatial scale, could they be governed by the mathematics of fractals?

Beginning in the late 1970s and 1980s with Benoit Mandelbrot’s pioneering work (Mandelbrot 1982), researchers investigated whether fractal geometry could provide a comprehensive language for natural shapes, terrains, and textures. In a seminal 1984 paper, Alex Pentland (Pentland 1984) showed that many natural surfaces—such as rolling hills, clouds, vegetation, and rough textures—behave statistically as fractional Brownian motion (fBm). Pentland demonstrated that the estimated fractal dimension of an image patch correlates strongly with human perception of surface “roughness” or texture complexity.

This resource page is designed for students interested in taking a historical and computational deep dive into the relationship between fractals, scale invariance, and image systems.

45.2 The Mathematical Connection

45.2.1 Scale Invariance and Fractal Dimension

A standard two-dimensional surface \(z(x, y)\) generated by fractional Brownian motion (fBm) has stationary increments with variance scaling as a power of distance:

\[ \mathbb{E}\left[ |z(x + \Delta x, y + \Delta y) - z(x, y)|^2 \right] \propto (\Delta x^2 + \Delta y^2)^H \]

where \(H \in (0, 1)\) is the Hurst parameter (or persistence parameter).

When mapped to the Fourier domain, an isotropic fBm process yields a radial power spectral density that follows a power law:

\[ P(f) \propto \frac{1}{f^\alpha} \]

The spectral exponent \(\alpha\) directly determines the topological space-filling behavior, quantified by the fractal (Hausdorff) dimension \(D\):

\[ D = D_T + 1 - H = 3 - \frac{\alpha - 1}{2} = \frac{7 - \alpha}{2} \]

or, for image luminance surfaces often parameterized directly against the exponent \(\eta \approx \alpha\):

\[ D = 3 - \frac{\eta}{2} \]

where \(D\) quantifies space-filling characteristics (with topological dimension 2 and embedding dimension 3, so \(2 \le D < 3\)).

Natural images typically exhibit estimated dimensions in the range:

\[ D \approx 2.2 \text{ -- } 2.6 \]

consistent with measured \(1/f^\alpha\) spectra where \(\alpha \approx 1.8\text{--}2.4\) (Pentland 1984; Zoran and Weiss 2009).

Under this mathematical framework, power-law spectra imply statistical scale invariance: analyzing fine details over small regions yields statistical properties equivalent to those obtained when analyzing coarse features over large fields of view.

45.3 Why Image Systems Engineering Diverged from Fractals

Despite the elegance of fractal models, mainstream image systems engineering and modern computer vision largely shifted toward other frameworks:

  1. Occlusion vs. Self-Similarity: Natural scenes are composed of distinct, opaque objects located at different depths. Fractals assume recursive self-similarity without boundaries. In contrast, the Dead Leaves model (Lee et al. 2001) demonstrates that sharp step discontinuities caused by object boundaries and occlusion naturally produce \(1/f^\alpha\) power spectra without requiring self-similar geometry.
  2. Finite Scale Bounds: Physical systems are strictly bounded. At large scales, perspective geometry and horizon limits break scale invariance; at small scales, optical diffraction limits, sensor pixel apertures, and atomic surface structures impose cutoffs.
  3. The Rise and Fall of Fractal Image Compression: In the 1990s, Iterated Function Systems (IFS) were proposed for image compression (e.g., Barnsley, Jacquin). While mathematically fascinating, finding contractive affine self-transformations across an image was computationally expensive, and the approach was rapidly surpassed by wavelet transforms (JPEG 2000), block DCTs, and modern learned codecs.
  4. Local Wavelet and Diffusion Priors: Modern visual models describe natural scenes using multiscale wavelet frames, steerable pyramids, Markov random fields, and score-based diffusion models that capture rich, non-Gaussian phase structures that fractals overlook.

45.4 Suggested Readings for a Deep Dive

Students interested in exploring the historical development and mathematical nuances of fractals in vision should consult:

  • Alex P. Pentland (1984): “Fractal-Based Description of Natural Scenes”, IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-6(6), pp. 661–674 (Pentland 1984). (The foundational vision paper connecting fBm to texture perception and 3D surface shape).
  • Benoit B. Mandelbrot (1982): “The Fractal Geometry of Nature”, W. H. Freeman and Company (Mandelbrot 1982).
  • Daniel Zoran and Yair Weiss (2009): “Scale Invariance and Noise in Natural Images”, IEEE International Conference on Computer Vision (ICCV) (Zoran and Weiss 2009).
  • Arnaud E. Jacquin (1992): “Image coding based on a fractal theory of iterated contractive image transformations”, IEEE Transactions on Image Processing, 1(1), pp. 18–30.
  • Antonio Torralba and Aude Oliva: Studies on natural scene statistics and scale invariance across scene categories (MIT Scene Statistics Notes).

45.5 Student Project Opportunities: Developing Novel ISETCam Code

Research & Code Opportunity with ISETCam

Currently, ISETCam does not include built-in code for generating or analyzing fractal textures and scenes.

If you are a student interested in exploring this topic for a term project or independent research, this presents an excellent opportunity for novel code development! You can collaborate directly with the AI assistant (Antigravity) to design, implement, and validate new ISETCam tools in MATLAB.

Here are concrete project directions you could explore:

  1. Synthetic Spectral Fractal Scene Generator (sceneCreate('fractal', ...)):
    • Implement 2D spectral synthesis of fractional Brownian surfaces with controllable spectral exponent \(\alpha\) (or Hurst parameter \(H\)).
    • Assign realistic surface reflectance spectra (e.g., from the Macbeth or natural surface databases in ISETCam) to generate full multispectral ISETCam scene objects.
  2. Impact of Optics and Sensor Sampling on Fractal Dimension:
    • How does optical blur (diffraction, lens defocus, chromatic aberration via oiCompute) distort the measured power-law spectrum of a fractal surface?
    • How does pixel sampling and pixel fill-factor in sensorCompute bias the estimated fractal dimension \(D\)?
    • Does sensor noise (Poisson shot noise and read noise) produce an apparent shift in the estimated Hurst parameter at high spatial frequencies?
  3. Fractals vs. Dead Leaves in Sensor Validation:
    • Compare how standard camera pipeline algorithms (demosaicing, noise reduction, sharpening) perform on synthetic fractal textures versus dead-leaves occlusion charts.

If you would like to undertake this, start an interactive session with the AI assistant to outline your algorithm and draft your first MATLAB/ISETCam script!

Return to Spatial Properties of Light Fields | Return to book home