AI Glossary / Computer vision and multimodal AI / Intermediate

Gabor transform

Definition

The Gabor transform is a time-frequency representation obtained by applying a Fourier transform to a signal viewed through a shifted Gaussian window. It shows which frequency components occur near different times. Related spatial constructions analyze local frequency content in images.

What it means

An ordinary Fourier transform describes the frequencies present across an entire signal, but does not directly show when a short event occurred. The Gabor transform adds localization by giving nearby samples more weight than distant ones during each analysis.

The window is moved across the signal, producing a sequence of local spectra. This helps distinguish, for example, a tone that occurs early from one that occurs later even when the overall frequency content is similar.

The term has some variation in mathematical usage, especially for discrete Gabor analysis and frame theory. Here it means the familiar Gaussian-windowed short-time Fourier transform. Broader Gabor systems may use other window choices or sampling constructions, so the convention should be stated.

How it works

For a window centered at time τ, the method multiplies the signal by a Gaussian centered at τ. It then takes the Fourier transform of that weighted segment. Repeating the process for different centers produces coefficients indexed by both time and frequency.

The coefficients generally contain magnitude and phase. A magnitude or power display is often called a spectrogram, but that display is not the full complex transform.

Window width creates a trade-off. A narrow window localizes events more tightly in time but gives less precise frequency discrimination. A broad window improves frequency discrimination while spreading abrupt changes across time. Discrete implementations also choose a hop size, frequency sampling and boundary treatment. These settings affect redundancy and reconstruction properties.

Three practical examples

1. Audio inspection: An analyst locates when a tone or vibration component appears, rather than seeing only its presence in a whole-recording spectrum.

2. Machine monitoring: A system describes changing frequency content in a sensor signal. Later analysis determines whether those changes are relevant to operating conditions.

3. Image texture analysis: A spatially localized frequency representation helps describe how patterns vary across an image, extending the windowed-analysis idea from time to position.

Think of it this way

Think of listening to a long recording through a moving, softly fading time window. At each position, you ask which notes are strongest nearby, then move on and repeat.

The analogy explains locality but not the full mathematics. The coefficients encode phase as well as strength, and the window's width changes the balance between knowing when something happened and distinguishing closely spaced frequencies.

Worked example

Imagine a recording that contains a 200 Hz tone during its first second and a 500 Hz tone during its second second. A whole-recording Fourier spectrum can show both frequencies but does not directly identify their order.

A Gaussian-windowed analysis centered well inside the first second should emphasize the 200 Hz component. A window centered well inside the second should emphasize 500 Hz. Near the transition, both may contribute according to the window width.

Now compare a narrow window with a broad one. The narrow window can locate the change more closely, while the broad window combines more samples and may distinguish nearby tones more clearly. Exact resolution depends on the chosen parameters; these two-tone observations illustrate the trade-off rather than specifying a universal window length.

Illustrative case study

Fictional case: Acorn Equipment analyses recordings from a small motor. A global spectrum shows two strong frequency components but does not reveal whether they occur together or during different operating stages. A local time-frequency view shows that one component appears mainly during startup.

The engineer aligns the analysis with the recorded operating sequence and tests different window widths to avoid confusing a short transition with a sustained condition. The team then evaluates whether the feature is useful for its monitoring objective. The case demonstrates interpretation of local frequency information, not a diagnosis or a measured reliability claim.

Common misconceptions and mistakes

A Gabor transform does not provide unlimited time and frequency precision simultaneously. Window choice imposes a fundamental localization trade-off.

A spectrogram image is not the complete transform. Squared magnitude omits phase, which can matter for reconstruction and further processing.

Using more closely spaced window centers does not automatically overcome the resolution limits of the chosen window. It can produce a denser and more redundant representation.

The Gabor transform and a single Gabor filter are related but not identical descriptions of a workflow. A transform collects a structured set of local frequency coefficients; one filter supplies a particular response.

Business use

The transform is useful when the timing or location of frequency content matters. It can support feature development for audio, vibration and image analysis, particularly when a single global spectrum hides changing behavior.

Teams should select window parameters using the events and frequency separations relevant to the application. Document units, sampling rate, window width and hop size. Compare features against labeled operating conditions, and avoid interpreting every visible band as an independently meaningful event.

People also ask

Is the Gabor transform a type of short-time Fourier transform?

Under the common convention used here, yes: it uses a Gaussian analysis window.

Can the original signal be reconstructed?

Suitable transform sampling, window choices and synthesis procedures can support reconstruction. A plotted magnitude-only image generally does not retain everything needed by a straightforward inverse.

Is the Gaussian window strictly limited to a finite interval?

Mathematically it extends indefinitely while decaying. Practical implementations truncate or otherwise approximate it.

Check your understanding

An analyst halves the hop size but keeps the same Gaussian window width. Has the fundamental time-frequency resolution trade-off disappeared?

Answer: No. The representation is sampled more densely in time, but the window still governs localization and frequency discrimination. More displayed columns are not equivalent to unlimited additional information.

Related terms

• Short-Time Fourier Transform
• Spectrogram
• Gabor filter
• Phase (Fourier)
• Gaussian window
• time-frequency analysis
• window function
• phase

Learn next

Analyze a signal with two tones that occur at different times, then one with two nearby simultaneous frequencies. Compare several Gaussian window widths. Inspect complex coefficients as well as magnitude displays, and study the conditions required for reconstruction before discarding phase.

More in Computer vision and multimodal AI

• Frame buffer - A frame buffer is memory used to hold image data for a frame, such as the pixels being rendered or prepared for display.
• Frame interpolation - Frame interpolation creates estimated images at times between existing video frames or related photographs.
• Full-reference metric - A full-reference metric evaluates the quality or similarity of an image or video by comparing it with a complete reference version.
• Galerkin approximation - A Galerkin approximation represents an unknown solution using a finite set of basis functions and chooses their coefficients so that the equation's residual...
• Gamma correction - Gamma correction is a nonlinear adjustment of image values used to account for, encode or modify the relationship between numerical values and light or...
• Gamma encoding - Gamma encoding is the use of a nonlinear transfer curve to represent linear-light image values as stored or transmitted code values.

Sources

Primary documentation: MathWorks, “From Fourier Analysis to Wavelet Analysis,” https://www.mathworks.com/help/wavelet/gs/from-fourier-analysis-to-wavelet-analysis.html ; and “Nonstationary Gabor Frames and the Constant-Q Transform,” https://uk.mathworks.com/help/wavelet/gs/non-stationary-gabor-frames.html . Relevant returned explanations of windowed Fourier analysis and Gabor terminology were reviewed. The two-tone example is an original conceptual illustration.

Previous: Gabor filter Next: Gain parameter