Does a stationary image have spatial frequency?

Yes. The variation is across position, so motion is unnecessary.

Do narrower stripes mean higher frequency?

Yes, when the unit of distance or visual angle is held fixed.

Can resizing change spatial frequency?

It changes the frequency expressed in output pixels; the physical or angular frequency also depends on how the result is displayed.

AI Glossary / Computer vision and multimodal AI / Beginner

Spatial frequency

Definition

Spatial frequency describes how rapidly a pattern varies across space, often measured as cycles per unit distance or per degree of visual angle. A cycle is one complete repetition of a pattern. High spatial frequencies correspond to finer variations; low spatial frequencies correspond to broader changes. The unit and viewing scale must be stated.

What it means

A wide set of alternating light and dark bands has fewer repetitions across a fixed distance than a tightly packed set. Their spatial frequencies differ even if their colors and contrast are identical.

The term is not limited to regular stripes. Fourier analysis represents more complex images as combinations of components with different spatial frequencies, orientations, phases, and amplitudes. This offers a way to discuss broad shading separately from fine texture, while recognizing that a sharp boundary involves multiple frequencies.

Spatial frequency is different from temporal frequency. One describes variation with position; the other describes variation with time. A stationary striped image has spatial variation without requiring motion. If it moves, the two kinds of variation can become related.

How it works

For a one-dimensional sinusoidal pattern, write I(x) = m + A cos(2πfx + φ). Here m is the mean level, A the amplitude, f the spatial frequency, and φ the phase. The period is 1/f in the same spatial units.

In two dimensions, a frequency vector specifies both the rate and direction of variation. The visible stripes run perpendicular to that vector. Changing phase shifts the pattern without changing its frequency.

Digital images often use cycles per pixel or cycles across the image width. Vision experiments commonly use cycles per degree of visual angle, which requires the display geometry and viewing distance. Physical imaging may use cycles per millimeter. These are different coordinate descriptions of a pattern, not interchangeable labels.

Sampling imposes limits. On a one-dimensional unit-spaced grid, the Nyquist frequency is 0.5 cycles per sample. Frequencies above that can masquerade as lower frequencies unless appropriately filtered before sampling.

Three practical examples

1. Camera evaluation: a team compares how well an imaging system preserves contrast in patterns with increasingly fine spatial detail under a defined test geometry.

2. Fabric photography: fine repeating threads can exceed the sensor's usable sampling capacity, creating false coarse patterns rather than a faithful record of the weave.

3. AI image preparation: a team resizes images for a classifier and checks whether downsampling removes small features or introduces aliases that differ from the deployment images.

Think of it this way

Think of musical pitch translated into a pattern on a page: closely spaced repetitions resemble a high pitch, and widely spaced repetitions resemble a low pitch.

The analogy changes the independent variable. Musical pitch concerns cycles over time, while spatial frequency concerns cycles over distance or angle. An image also has orientation and location, so the comparison is helpful but incomplete.

Worked example

Suppose a sinusoidal pattern repeats every eight pixels. Its spatial frequency is 1/8 = 0.125 cycles per pixel. Across 256 pixels, that corresponds to 32 cycles, ignoring the endpoint-counting details of a particular finite crop.

If the same displayed region subtends four degrees at the viewer's eye, those 32 cycles correspond to eight cycles per degree. If its angular width changes, cycles per degree changes even though the image file still contains 32 cycles. This is why a pixel-only description cannot fully specify a visual stimulus.

Now sample a one-dimensional cosine of frequency 0.75 cycles per pixel at integer positions. For integer n, cos(2π × 0.75n) has the same samples as cos(2π × 0.25n). The sampled values cannot tell those two frequencies apart. The higher-frequency pattern has aliased to a lower one.

This example assumes an ideal uniform grid. Real pixel integration, optics, and processing complicate observed aliases. A coarse pattern appearing after sampling does not prove the scene contained that pattern.

A frequency of exactly 0.5 cycles per pixel is a boundary case whose sampled appearance depends on phase. Practical imaging needs margin and suitable filtering rather than relying on ideal equality at the limit.

Illustrative case study

Fictional case: an online textile shop finds that one shirt photograph develops broad colored ripples when reduced for thumbnails. The original fabric contains a fine repeating weave, and the artifact changes when the image is resized differently.

The imaging team compares controlled reductions with appropriate prefiltering and inspects both still thumbnails and zoom transitions. It also reviews capture distance and lens settings, since some aliasing may already have occurred before the file reached the website pipeline.

The team distinguishes fabric texture from sampling artifacts. A later filter cannot guarantee recovery of ambiguously sampled detail.

Common misconceptions and mistakes

Do not equate high spatial frequency with useful detail in every case. Noise and artifacts can also contribute high-frequency energy.

Do not describe an edge as one single frequency. A sharp transition requires a range of components, so removing high frequencies can soften it.

Avoid comparing cycles per pixel with cycles per degree without a scale conversion. Also distinguish pattern frequency from contrast: two gratings can have the same frequency but different amplitudes, and their visibility can differ. A frequency-domain representation does not by itself tell you which structures are meaningful to a person or model.

Business use

Spatial-frequency reasoning helps teams choose capture resolution, resizing methods, display tests, and image-quality criteria. More pixels and stronger sharpening are not interchangeable improvements.

Identify the smallest relevant structures at the final processing scale. Check whether capture and resampling preserve them. Evaluate the actual task: reproducing a grating does not fully determine recognition of products or defects.

Check your understanding

A pattern has one complete cycle every ten pixels. What is its frequency in cycles per pixel?

Answer: 0.1 cycles per pixel. To express it in cycles per degree, you also need to know how many pixels occupy a degree of visual angle under the viewing setup.

Related terms

Fourier transform
Sampling
Nyquist frequency
Aliasing (factorial experiments)
Phase (Fourier)
aliasing
spatial period
phase
modulation transfer function
contrast sensitivity

Learn next

Compare gratings while changing only one property at a time: frequency, orientation, phase, or contrast. Then examine how downsampling changes their appearance with and without prefiltering, keeping the input and output coordinate units explicit.

Sources

NYU, Spatial Frequency Channels, primary teaching outline: https://www.cns.nyu.edu/~david/courses/perception/outlines/outline-channels.html NYU, Perception Lecture Notes, Spatial Frequency Channels: https://www.cns.nyu.edu/~msl/courses/0022X/lecturenotes/channels/channels.html Teaching material was read for frequency representation and visual angle. Calculations and the case are original.

Previous: Spatial embedding Next: Spatial intelligence (artificial intelligence)