Skip to content

Fractals and Multiscale Texture Analysis

Published by Ka Ming Lui on September 1, 2026

Texture is not defined at only one size. A surface may appear smooth when viewed from far away and highly irregular when examined more closely. Fractal and multiscale analysis describe how structure changes as the measuring scale changes.

The central question is not only what pattern is present, but also how the amount of measured detail changes across scale.
Quick Navigation
1. Why Scale Matters 2. Self-Similarity 3. Types of Fractal Structure 4. Deterministic Fractals 5. Fractal Dimension 6. Measuring Instrument 7. Box Counting 8. Log-Log Slope 9. Fractal Brownian Function 10. Hurst Exponent 11. Spectral Scaling 12. Real-World Examples 13. Local vs. Global 14. Multiscale Features 15. Practical Limits 16. Interactive Visualizer 17. Summary

1. Why Scale Matters

The amount of visible detail depends on the size of the observation or measuring scale. Coarser measurements suppress small variations, while finer measurements preserve increasingly small structure.

The Same Profile Viewed at Five Effective Scales

Very Coarse

strong averaging

Only the broadest trend remains.

Coarse

coarse averaging

Large-scale rises and falls are visible.

Intermediate

moderate averaging

Mid-scale peaks and valleys begin to appear.

Fine

light averaging

Most local changes are preserved.

Very Fine

original sampled detail

Fine-scale irregularity contributes fully.

A multiscale descriptor does not describe the image at only one fixed scale. It records how a measurement changes from very coarse to very fine scales.

2. Self-Similarity

A self-similar structure contains organization that reappears after a change of scale. In an exact deterministic fractal, a selected part can reproduce the same organization after rescaling. Natural objects more often show statistical self-similarity over only a finite range of scales.

A Selected Part Repeats the Organization of the Whole

Whole Pattern

The orange outline selects one part whose width is one third of the full pattern.

Selected Region Enlarged

Enlargement reveals the same eight-around-one-hole organization at the next finer scale.

Exact self-similarity repeats a rule exactly. Statistical self-similarity preserves related roughness or organization rather than identical copies.
Real-world structures are normally fractal-like over a limited range of scale rather than from infinitely large to infinitely small sizes.

3. Different Types of Fractal Structure

The word fractal can describe more than one kind of scaling behavior. Two especially useful cases for image analysis are a deterministic fractal and a fractal Brownian function.

Two Scaling Models Used on This Page
Deterministic

Deterministic Fractal

one stage next stage

One fixed segment-replacement rule is applied again at every stage. The scaling is exact and predictable.

Scaling relation 1 = nλD
Statistical

Fractal Brownian Function

The exact waveform is random, but its roughness follows a statistical scaling law described by the Hurst exponent \(H\).

Fractal dimension D = 2 − H
Fractal Curve and Fractal Surface Describe Different Geometric Settings

These are not separate fractal-generation rules. They describe what is produced when a fractal Brownian model is defined on a one-dimensional or two-dimensional domain.

1-D Fractal Brownian Function → Fractal Curve

position function value
Fractal dimension D = 2 − H

2-D Height Function → Rough Fractal Surface

x y height 2-D position domain (x, y) with height shown vertically
A two-dimensional height function assigns a height to each position (x, y). In this plot, both vertical displacement and surface shading represent that height, so peaks and valleys are visible directly.
Fractal dimension D = 3 − H
A deterministic fractal demonstrates exact scale replacement. A fractal Brownian function describes statistical roughness. Whether that statistical model forms a fractal curve or a fractal surface depends on the dimension of its domain.

4. Deterministic Fractals

A deterministic fractal can be generated by repeatedly applying the same replacement rule. The example below starts with one line segment. Every segment is replaced by four smaller segments, each scaled by one third.

One Segment-Replacement Rule, Applied Repeatedly
1 segment

Stage 0

Begin with one unit segment.

4 segments

Stage 1

Replace the original segment with four smaller segments.

16 segments

Stage 2

Apply the same replacement rule to all four segments.

\(n=4\) four smaller segments replace each previous segment
\(\lambda=1/3\) each new segment has one-third the previous length

5. Fractal Dimension

Fractal dimension \(D\) describes how rapidly measured detail or occupied structure increases as the measuring scale becomes finer. For an exactly self-similar deterministic fractal with \(n\) copies, each scaled by \(\lambda\), the unit measure obeys

\[ 1=n\lambda^D. \]

Solving for \(D\) gives

\[ D = -\frac{\log n}{\log \lambda} = \frac{\log n}{\log(1/\lambda)}. \]
Fractal Dimension Does Not Have to Be an Integer
\(D=1\)

Ordinary Line

An ideal smooth line has integer dimension \(1\).

\(D\approx1.262\)

Deterministic Fractal Curve

For \(n=4\) and \(\lambda=1/3\), \(D=\log 4/\log 3\approx1.262\).

\(D=2\)

Filled Plane Region

A filled plane region has integer dimension \(2\).

Non-integer dimension is allowed. A fractal curve can have \(1<D<2\): it is more space-filling than a smooth line without becoming a filled plane.
A fractal Brownian function gives another example. If \(H=0.3\), then \(D=2-H=1.7\), a non-integer dimension associated with a relatively rough fractal structure.

6. Measuring with a Measuring Instrument

Let \(\lambda\) be the length of the measuring instrument at a particular scale. A large \(\lambda\) skips small bends. As \(\lambda\) becomes smaller, more placements are needed to follow the same unit measure.

The Same Irregular Curve Measured at Five Scales

Very Large \(\lambda\)

Only the broad route is measured.

2 placements

Large \(\lambda\)

A few major bends begin to contribute.

4 placements

Medium \(\lambda\)

Intermediate structure is followed.

5 placements

Small \(\lambda\)

Most local bends are included.

8 placements

Very Small \(\lambda\)

The measuring instrument follows the finest sampled detail.

14 placements
As \(\lambda\) decreases, the number \(n\) of measuring-instrument placements generally increases. Fractal dimension describes the scaling relationship between \(n\) and \(\lambda\).

7. Box Counting

For a sampled image, box counting uses a square box as the measuring instrument. At each box side length \(\epsilon\), count every box that contains at least one foreground point:

\[ N(\epsilon) = \text{number of occupied boxes at scale }\epsilon. \]
One Pattern, Five Box Sizes

Very Coarse

\(\epsilon=1\)
\(N(\epsilon)=1\)

Coarse

\(\epsilon=0.5\)
\(N(\epsilon)=3\)

Intermediate

\(\epsilon=0.25\)
\(N(\epsilon)=9\)

Fine

\(\epsilon=0.125\)
\(N(\epsilon)=27\)

Very Fine

\(\epsilon=0.0625\)
\(N(\epsilon)=81\)
1. Choose \(\epsilon\)Select a box side length.
2. Overlay the gridCover the same pattern.
3. Count occupied boxesRecord \(N(\epsilon)\).
4. Repeat across scaleUse several coarse-to-fine values.
In this example, decreasing the box side length gives \(N(\epsilon)=1,3,9,27,81\). The repeated factor of three indicates a stable scaling pattern across these five levels.

8. The Log-Log Slope

If the occupied-box count follows a power law over a useful scale range,

\[ N(\epsilon)\propto\epsilon^{-D}. \]

Taking logarithms gives a straight-line relationship:

\[ \log N(\epsilon) = D\log\!\left(\frac{1}{\epsilon}\right)+C. \]
Five Scales Form a Linear Trend on Log-Log Axes
0 0.69 1.39 2.08 2.77 log(1/ε) log N(ε)
Scale range
ε = 1, 1/2, 1/4, 1/8, 1/16
Estimated fractal dimension
D = log 3 / log 2 ≈ 1.585
The estimate should come from a range where the log-log points are approximately linear. Curvature can indicate that one fractal dimension does not describe every available scale.

9. Fractal Brownian Function

Natural textures are usually not exact deterministic fractals. A fractal Brownian function provides a statistical model in which the size of an increment depends on the separation \(\Delta\):

\[ E\!\left\{|X(t+\Delta)-X(t)|^2\right\} \propto |\Delta|^{2H}. \]
The Same Statistical Model Can Produce Different Roughness Levels

Rough Fractal Structure

lower \(H\), for example \(H=0.2\)

Strong fine-scale variation makes the graph more jagged.

Intermediate Structure

middle \(H\), for example \(H=0.5\)

Both broad and local changes contribute visibly.

Smooth Fractal Structure

higher \(H\), for example \(H=0.8\)

Nearby values change more gradually.

10. Hurst Exponent, \(H\), and Fractal Dimension, \(D\)

The Hurst exponent \(H\) controls the statistical roughness of a fractal Brownian function. For \(0<H<1\), lower \(H\) corresponds to rougher behavior and higher \(H\) to smoother behavior.

1-D Fractal Brownian Function

D = 2 − H

Here \(D\) is the fractal dimension of the function’s graph. It lies between an ordinary line and a filled two-dimensional region.

2-D Fractal Brownian Surface

D = 3 − H

Here \(D\) is the fractal dimension of the rough surface. The surface itself can be represented by a 2-D height map when a perspective drawing is not useful.

Example: for a 1-D fractal Brownian function, \(H=0.3\) gives \(D=2-0.3=1.7\). This is both a non-integer fractal dimension and a relatively rough fractal structure.
The relationship between \(H\) and \(D\) depends on whether the modeled object is a 1-D function or a 2-D surface. State the geometric setting whenever a fractal dimension is reported.

11. Spectral Scaling

Scaling behavior can also be described in the frequency domain. For a one-dimensional fractal Brownian function,

\[ S(f)\propto f^{-(2H+1)}. \]
A Power Law Becomes a Straight Trend on Log-Log Axes
log f log S(f)
Power law S(f) ∝ f−(2H + 1)
Log form log S(f) = −(2H + 1) log f + C
Line slope m = −(2H + 1)
The equations and fitted slope are kept outside the graph. The plot itself only needs to show the measured spectral trend clearly.

12. Fractal-Like Structure in the Real World

Exact mathematical fractals are ideal constructions, but many natural shapes show related organization or roughness over a finite range of scales. These are better described as fractal-like or statistically self-similar rather than exact copies at every scale.

Examples of Scale-Dependent Structure

Coastline-Like Boundary

A shorter measuring instrument reveals additional bends, so the measured boundary length can depend on scale.

Tree and Fern Branching

Large branches divide into smaller branches, producing related organization across several biological scales.

Cloud and Terrain Roughness

Cloud boundary
Terrain profile

Cloud boundaries and terrain profiles can show statistical roughness over several observation scales. Fractal Brownian functions or surfaces provide one way to model that scaling behavior over a finite range.

Rough Material Texture

Fine and coarse surface irregularities can coexist, making a multiscale description more informative than one fixed-scale statistic.

These examples are not claimed to be ideal infinite fractals. The useful question is whether a stable scaling relationship exists over a meaningful range of real measurement scales.

13. Local vs. Global Fractal Description

A whole image may contain several kinds of texture. One global fractal dimension can therefore hide important regional differences. Local windows allow the estimated scaling behavior to change from one image region to another.

Different Regions Can Have Different Scaling Behavior

Window A

A local \(D\) or \(H\) estimate describes one texture region.

Window B

A different scaling estimate can describe another region.

Local fractal features can be displayed as texture maps, allowing spatial changes in roughness or scaling behavior to become visible.

14. Multiscale Texture Features

Fractal descriptors do not have to be used alone. A texture description can combine fractal dimension, Hurst exponent, and other measurements at several scales.

\[ \mathbf{x} = [ D, H, E_1, E_2, \ldots ]. \]

Here, \(D\) denotes the chosen fractal-dimension estimate, such as the value obtained from box counting. The measurement method is named in the surrounding text rather than placed in a subscript.

Texture image or local region
↓
Measure structure at several scales
↓
Estimate scaling features such as \(D\) or \(H\)
↓
Combine with other texture descriptors when useful
↓
Compare, cluster, or classify textures

15. Practical Limits of Fractal Texture Analysis

Fractal dimension is useful only when the measuring procedure and scale range are interpreted carefully.

Finite Image Size Very coarse scales provide only a few boxes or measuring-instrument placements.
Pixel Resolution The finest meaningful scale cannot be smaller than the available sampling resolution.
Scale Range A power law may hold over only part of the available range.
Thresholding Binary box counting can change when the foreground definition changes.
Grid Alignment Shifting the box grid can slightly change occupied-box counts in a finite image.
Noise Fine-scale noise can artificially increase apparent roughness.
A single value of \(D\) should not be treated as proof that a real object is an ideal fractal. The stronger evidence is a stable scaling relationship over a meaningful range of measurement scales.

16. Interactive Fractal and Multiscale Visualizer

The visualizer below provides three complementary ways to explore exact and statistical fractal structure.

Build Your Fractal

Draw a generator, change the scale factor, choose more recursion depth, and watch the deterministic fractal expand automatically. The visualizer reports \(n\), \(\lambda\), \(D\), and whether the resulting dimension is integer or non-integer.

Box Counting

Edit a binary pattern, move through five box sizes from very coarse to very fine, inspect occupied boxes, and estimate fractal dimension from the log-log slope.

Fractal Brownian Function

Change \(H\), vary the number of scales in the generated structure, compare rough and smooth fractal structure, and inspect increment scaling together with the estimated \(D\).

17. Summary

Concept Main Idea
Scale Visible structure changes from coarse to fine measurements.
Self-similarity Related organization reappears after a change of scale.
Deterministic fractal The same segment-replacement or generator rule is repeated exactly.
Fractal Brownian function Statistical roughness follows a scale-dependent law rather than exact copying.
Measuring instrument \(\lambda\) is the measuring-instrument length at a particular scale.
Fractal dimension \(D\) can be integer or non-integer and describes how detail grows as scale becomes finer.
Deterministic scaling For exact self-similarity, \(1=n\lambda^D\).
Box counting Count occupied boxes \(N(\epsilon)\) across several box sizes.
Log-log slope The slope of \(\log N\) versus \(\log(1/\epsilon)\) estimates \(D\).
Hurst exponent Lower \(H\) gives rougher structure; higher \(H\) gives smoother structure.
1-D graph dimension For a 1-D fractal Brownian function, \(D=2-H\).
2-D surface dimension For a 2-D fractal Brownian surface, \(D=3-H\).
Real-world fractal-like structure Natural objects often show approximate or statistical scaling over a finite range.
Local analysis Different image regions can have different scaling behavior.
Fractal and Multiscale Texture Visualizer

Explore how measured structure changes with scale.

Draw a small generator pattern. At each depth, every active cell is replaced by a scaled copy of that same generator, creating a self-similar fractal from your shape.
Active copies n—
Scale λ—
Similarity D—
Dimension type—
Current depth—
Logical cells / side—
Your Generator
Tap cells to turn them on or off. Blue cells are copied at every scale.
If the generator has n active cells in an s × s grid, the measuring scale is λ = 1/s. For exact self-similar replacement, 1 = nλD, so D = log(n) / log(s). Very deep patterns are automatically resampled so they remain visible on narrow screens.
Generated Self-Similar Pattern
Edit the generator or press Generate.
Click cells to edit the binary pattern. Change the box size to see which boxes are occupied, then process several scales to estimate the log-log slope.
ε = 1
Binary Pattern and Box Overlay
foreground occupied box
16 × 16 binary image
Box-Counting Measurements
Estimated D—
R²—
Occupied Now—
BoxεN(ε)log(1/ε)log N
Choose a scale or analyze the pattern.
Move the Hurst exponent and change the number of scales. Lower H produces a rougher fractal structure; higher H produces a smoother fractal structure.
0.50
Seed 1337
Chosen H0.50
Estimated H—
Graph D = 2 − H—
Roughness—
Scale levels7
2-D Roughness Preview
The preview combines several spatial scales. Increase the scale-level control to add finer structure.
1-D Fractal Brownian Function
Increment Scaling
Scaling relation log M(Δ) ≈ m log Δ + C
Fitted slope m —
Estimated H = m / 2 —
The graph shows the measured points and fitted trend; the fitted scaling values are summarized below.

Published inDigital Image Processing

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *