Does it work on any motion size?

No. The phase-to-motion approximation and representation impose limits; large movement can create artifacts.

Is it the same as slowing a video down?

No. Slow motion changes playback timing, whereas magnification changes the apparent amplitude of selected variations.

Can it measure vibration automatically?

It can support analysis, but reliable physical measurement requires calibration, suitable capture and a validated measurement procedure.

AI Glossary / Computer vision and multimodal AI / Intermediate

Phase-based magnification

Definition

Phase-based magnification is a video-processing technique that makes small motions more visible by modifying temporal changes in the local phase of image representations. It amplifies selected motion-related variations and reconstructs a video in which those movements appear larger.

What it means

Some movements are present in a recording but are too small to notice easily. A vibrating object, for example, may move only a fraction of a pixel between frames. Phase-based processing represents local image structures at different spatial scales and orientations, then studies how their phase changes over time. It can magnify subtle movement without explicitly following every object as a tracked trajectory. The output is a processed visualization, not an untouched recording. Its apparent motion depends on filter settings, magnification strength and the limits of the input footage.

How it works

The method decomposes frames using complex-valued spatial filters, commonly a complex steerable pyramid. For each component, it separates amplitude from phase and filters phase variations over time to select a frequency band. It scales the selected variations and reconstructs the frames from the modified representation.
For sufficiently small movements, local phase changes relate to displacement, making this manipulation useful for motion amplification. The relationship has limits set by spatial scale and motion size. Large movements, low-amplitude regions, occlusions and phase wrapping can introduce artifacts. Camera motion or lighting changes can contaminate the signal, so capture conditions and preprocessing matter. The amplified clip should be checked against the original.

Three practical examples

An engineering demonstration makes a small periodic vibration of a test structure easier to see.
A teaching video illustrates subtle motion at a selected frequency band.
A laboratory visualization highlights small surface movements for further analysis using an independently defined measurement method.

Think of it this way

Think of turning up a selected movement channel in a video, somewhat as an audio equalizer emphasizes a frequency band. This conveys selection and amplification. The comparison has limits: the method operates on local spatial phases, and increasing the setting can distort the image rather than merely revealing an unchanged event more clearly.

Worked example

Suppose a controlled one-dimensional pattern has a spatial wavelength of 20 pixels. A displacement of 0.2 pixels corresponds in magnitude to a phase change of 2π × 0.2/20, approximately 0.0628 radians. If the selected phase variation is scaled to five times its original value, the ideal small-motion model predicts an apparent displacement of 1 pixel.
This is an illustrative sinusoidal calculation, not a guarantee for arbitrary video. Some software defines its parameter as an added amplification rather than total gain, so a displayed value of five may not mean a fivefold output. Spatial scale, temporal filtering and the validity of the local-motion approximation must also be considered.

Illustrative case study

In a fictional equipment workshop, a team records a small vibrating panel using a fixed camera and stable lighting. It creates a magnified preview around a selected temporal band, then compares it with the original footage. A broad movement across the entire frame prompts a check of camera stability before interpreting the panel's behavior. The team uses the visualization to guide further inspection, while any quantitative vibration claim relies on a calibrated analysis. This case illustrates the distinction between making a pattern visible and establishing its physical magnitude or cause.

Common misconceptions and mistakes

A dramatic amplified movement does not mean the object moved that far in the original scene. Magnification does not create unlimited spatial or temporal resolution. A selected frequency band can also contain camera shake, illumination effects or unrelated motion. Excessive gain can produce halos, tearing or implausible movement. Another mistake is treating the processed video as a diagnosis of a mechanical problem. The visualization may support investigation, but identifying a cause requires the relevant engineering evidence and measurements.

Business use

Businesses considering this technique should define whether the goal is communication, inspection support or calibrated measurement. Preserve original footage and record the processing settings. Test stable and moving-camera conditions, repeated captures and known-motion examples. Use frame rates appropriate to the motion frequencies of interest, since temporal aliasing can misrepresent a signal before processing begins. Review both selected regions and the wider scene to avoid interpreting a global recording artifact as a local physical effect.

Check your understanding

A processed clip shows a large oscillation after phase magnification. Can that displayed amplitude be reported directly as the object's original displacement?

Answer: No. It reflects amplification and processing settings. Recovering physical displacement requires the original data, the processing model and appropriate calibration.

Related terms

Learn next

Learn Fourier phase and temporal sampling next. Then inspect a known-motion example at several magnification settings, noting when useful visibility gives way to artifacts. Compare the processed sequence with its source frames throughout.

Sources

Wadhwa, Rubinstein, Durand and Freeman, Phase-Based Video Motion Processing: https://people.csail.mit.edu/billf/publications/Phase-Based_Video.pdf MIT research project collection on video magnification: https://people.csail.mit.edu/mrub/vidmag/ Primary research search excerpts consulted; full source and editorial review remain pending.

Previous: Pessimistic description Next: Phase correlation