Is phase measured in pixels?

Phase itself is an angle. Converting a phase change to displacement requires the associated frequency and a suitable model.

Does a shift change magnitude?

An ideal circular translation preserves the discrete Fourier magnitude, but cropping, newly exposed borders and other practical changes can alter it.

What does phase wrapping mean?

Angles differing by a full turn describe the same phase, so a displayed angle may jump at the chosen range boundary.

AI Glossary / Computer vision and multimodal AI / Intermediate

Phase (Fourier)

Definition

Fourier phase describes the positional alignment of a sinusoidal component in a Fourier representation. Together with that component's magnitude and frequency, it determines how the component contributes to the reconstructed signal or image.

What it means

A Fourier transform represents a signal through oscillating components. Magnitude describes how strongly each component contributes; phase describes where its oscillation sits relative to the coordinate origin and the transform convention. In an image, relationships among phases help determine the positions and arrangements of structures. Two images can therefore share a magnitude spectrum yet look different. Phase is an angle, commonly expressed in radians. It is not simply a color or brightness value, and the phase of an individual frequency should not be interpreted as a complete description of one visible object.

How it works

A complex Fourier coefficient can be written as a magnitude multiplied by a complex exponential carrying its phase. Software often obtains the phase using an angle function applied to the coefficient. The inverse transform combines all components with their magnitudes and phases to reconstruct the signal.
Translating a signal changes Fourier phases in a frequency-dependent way while preserving magnitudes for an ideal circular shift of discrete data. This property supports alignment methods such as phase correlation. Phase is defined modulo a full turn, so displayed values may jump between equivalent angle representations. Where a coefficient's magnitude is zero, its phase carries no meaningful contribution; at very small magnitudes, noise can make phase unstable.

Three practical examples

An image-registration tool uses phase relationships to estimate a translation between two images.
A signal-processing lesson swaps magnitude and phase information to demonstrate their different roles in image structure.
A video-processing method tracks changes in local phase to study small movements across frames.

Think of it this way

Think of musicians playing notes of different pitches. The loudness of each note resembles magnitude, while the timing of each repeating waveform resembles phase. Combining the notes depends on both. This analogy explains relative alignment, but image Fourier components extend across space rather than forming isolated musical events, and an image's meaning is not carried by any single component.

Worked example

Consider f(x) = cos(2πx/8), a cosine with a period of 8 samples. A shift two samples to the right gives g(x) = f(x − 2) = cos(2πx/8 − π/2). The frequency and amplitude stay the same, while the phase changes by −π/2 under this expression's convention.
For a component with a different frequency, the same two-sample shift produces a different phase change. That coordinated pattern across frequencies can encode a translation. The sign depends on the chosen shift and transform conventions, so an implementation should test a known displacement rather than assume that every library uses the same displayed sign.

Illustrative case study

In a fictional inspection project, two photographs of the same printed target differ mainly by a small camera shift. A developer compares Fourier magnitudes and finds them similar, then uses a phase-correlation method to estimate the translation. The team checks the estimate against visible markers and excludes images with major perspective changes from this simple translation test. The example shows how phase can support alignment while also requiring a suitable motion model. It does not claim that phase alone resolves deformation, occlusion or every lighting change.

Common misconceptions and mistakes

Phase is not unimportant merely because a magnitude spectrum is easier to display. Nor does phase alone always preserve all useful image information. Magnitude and phase work together, and demonstrations that retain recognizable structure under one manipulation do not establish a universal rule for every image. Wrapped phase jumps need not indicate abrupt physical motion. Another mistake is trusting angles at frequencies with almost no signal energy. Those values may be numerically available but unreliable for the intended estimate.

Business use

For registration, reconstruction or motion-analysis tools, phase is usually an internal representation rather than a user-facing feature. Teams should test coordinate conventions, padding, windowing and expected transformations. Evaluate alignment on images with repeated patterns, weak texture and changing borders, since ambiguous or contaminated phase relationships can produce misleading estimates. Report the resulting displacement or processing effect with appropriate evidence, rather than presenting a complicated phase image as proof that the output is correct.

Check your understanding

A cosine has an 8-sample period and is shifted right by 2 samples. What phase offset appears in cos(2πx/8 + phase)?

Answer: The offset is −π/2, because replacing x with x − 2 subtracts 2π × 2/8 from the angle.

Related terms

Fourier transform
Phase correlation
Image registration
Spatial frequency
magnitude spectrum
complex number
sinusoid
phase wrapping

Learn next

Learn complex Fourier coefficients and the shift property next. Shift a small periodic signal by known amounts and compare magnitude and phase changes, then repeat with a cropped image to see why boundary handling matters.

Sources

NumPy, official Fourier-transform documentation: https://numpy.org/doc/2.2/reference/routines.fft.html NumPy, angle function documentation: https://numpy.org/doc/2.0/reference/generated/numpy.angle.html Primary documentation search excerpts consulted; full source and editorial review remain pending.

Previous: Phase dispersion minimization Next: Phase of Spectrum