The paper proposes fractal Brownian functions as a unified, scale‑invariant model for natural 3‑D surfaces such as mountains, trees, clouds, and textures. It argues that fractal dimension is a necessary descriptor because standard Euclidean measures fail to capture the self‑similar roughness of natural shapes. The author shows how the imaging process maps a fractal surface to a fractal intensity field, enabling the inference of surface fractal dimension directly from image data via Fourier or co‑occurrence statistics. Empirical studies on a variety of natural and synthetic textures demonstrate that most regions (≈77 %) are well approximated by a fractal Brownian model, and that the measured fractal dimension is stable across scale and linear intensity transforms. The paper then applies the model to texture segmentation, classification, shape‑from‑texture, and shading, illustrating that fractal‑based segmentation outperforms traditional co‑occurrence and correlation methods. A “constrained‑chance” representation is introduced: a random‑generator‑based shape generator whose output is limited by image‑derived constraints (primarily fractal dimension and outline). An example of reconstructing a mountain peak from a single image shows that a compact 3‑D description can be obtained. The work concludes that fractal‑based, constrained‑chance descriptions provide a compact, scalable, and physically grounded way to represent both individual natural objects and whole classes of shapes, and that measuring fractal dimension from images can aid shape‑from‑texture and shading analysis.
Topics
Fractal Brownian Functions as a Universal Model for Natural Surfaces
Introduces fractal Brownian functions, defines fractal dimension (D = 2‑H), and explains why fractal dimension is essential for consistent measurement of natural shapes.
Mapping Fractal Surfaces to Image Intensity Fields
Analyzes how the imaging process transforms a fractal surface into a fractal intensity surface, establishing a logarithmic relationship between surface and image fractal dimensions.
Empirical Validation of the Fractal Model on Real Textures
Presents experiments on natural and Brodatz textures, showing that 77 % of regions fit a fractal Brownian model, with only 8 % poor fits due to saturation or non‑fractal structure.
Fractal Dimension as a Scale‑Invariant Texture Descriptor
Describes methods (Fourier power spectrum, co‑occurrence statistics) for estimating fractal dimension from image patches and demonstrates stability across 4:1–8:1 scale changes.
Texture Segmentation via Fractal Dimension Histograms
Shows how local fractal dimension histograms can be thresholded to segment images (e.g., water vs. land, sky vs. desert) and how boundaries manifest as sub‑topological dimensions.
Comparative Performance of Fractal Segmentation
Compares a two‑parameter fractal segmenter (84.4 % accuracy) against correlation, co‑occurrence, and texture‑energy methods, highlighting its superior simplicity and accuracy.
Linking Fractal Model to Classical Texture Measures
Explains the theoretical connection between fractal dimension, Fourier spectral fall‑off, and co‑occurrence statistics, suggesting a unifying framework for texture analysis.
Shape‑From‑Texture Using Fractal Anisotropy and Perspective Gradients
Describes how differences in fractal dimension across x and y reveal anisotropy, and how gradual increases in measured dimension with distance provide perspective texture gradients for surface orientation estimation.