Phase correlation

Phase correlation is a frequency-domain method for estimating the relative displacement between two signals or images. It derives translational offset from the normalized phase difference between their Fourier transforms, rather than from direct comparison of their sample amplitudes. The method is widely associated with image registration, although the same mathematical construction applies to one-dimensional signals and higher-dimensional sampled fields.

For two images related only by translation, phase correlation produces an impulse-like response whose location represents the displacement. Because normalization removes the magnitude of each Fourier coefficient, the response depends principally on phase consistency across spatial frequencies. This property distinguishes phase correlation from unnormalized cross-correlation, in which high-energy frequency components can dominate the resulting similarity surface.

Mathematical formulation

Let (f(\mathbf{x})) and (g(\mathbf{x})) denote two (d)-dimensional signals, with (g) translated from (f) by a displacement vector (\boldsymbol{\Delta}):

[ g(\mathbf{x}) = f(\mathbf{x}-\boldsymbol{\Delta}). ]

Under the Fourier shift theorem, their transforms (F(\mathbf{k})) and (G(\mathbf{k})) satisfy

[ G(\mathbf{k})

e^{-i2\pi \mathbf{k}\cdot\boldsymbol{\Delta}} F(\mathbf{k}). ]

The normalized cross-power spectrum is defined, under one common sign convention, by

[ R(\mathbf{k})

\frac{ F^{}(\mathbf{k})G(\mathbf{k}) }{ \left|F^{}(\mathbf{k})G(\mathbf{k})\right| }, ]

where the asterisk denotes complex conjugation. For every frequency at which the denominator is nonzero, substitution of the translation relation gives

[ R(\mathbf{k})

e^{-i2\pi \mathbf{k}\cdot\boldsymbol{\Delta}}. ]

The inverse Fourier transform of this spectrum is an impulse centered at the displacement:

[ r(\mathbf{x})

\mathcal{F}^{-1}{R(\mathbf{k})}

\delta(\mathbf{x}-\boldsymbol{\Delta}), ]

where (\delta) denotes the Dirac delta distribution. In a sampled system, the continuous impulse becomes a discrete correlation peak, and its location is interpreted modulo the dimensions of the sampling lattice. Definitions using (F(\mathbf{k})G^{*}(\mathbf{k})) instead produce the opposite displacement sign without altering the underlying estimate.

The normalization discards the product magnitude (\lvert F^{*}G\rvert), retaining only the relative phase of corresponding coefficients. Frequencies with negligible magnitude contain poorly determined phase, so practical formulations replace the denominator by a regularized expression or suppress coefficients whose magnitudes approach numerical precision. These modifications change the weighting of spectral evidence without changing the translation theorem on which the estimator is based.

Historical development

C. D. Kuglin and D. C. Hines introduced the phase-correlation formulation for image alignment in 1975, connecting normalized cross-power spectra with displacement estimation through the Fourier shift theorem. Their treatment established the characteristic inverse-transform peak as a representation of relative translation and placed the method within the emerging field of digital image processing.

Later work integrated phase correlation with statistical models of additive noise, finite sampling, and nonperiodic image boundaries. These developments replaced the idealized impulse model with an analysis of peak width, sidelobe structure, estimator variance, and spectral coherence. The resulting literature treats phase correlation as both a deterministic transform method and an estimator whose behavior depends on image content and observation conditions.

Discrete and finite-domain behavior

For arrays of size (N_1\times\cdots\times N_d), the discrete Fourier transform represents translation as a phase ramp only under the transform's periodic boundary model. Consequently, the inverse normalized spectrum describes circular correlation, and a displacement outside the principal coordinate interval appears as its wrapped equivalent. This ambiguity is intrinsic to finite discrete spectra rather than a separate numerical error.

Ordinary images rarely satisfy periodic boundary conditions because opposite edges generally contain unrelated intensities. Their discontinuity at the implied periodic boundary introduces broad spectral components and produces sidelobes in the correlation surface. A window function reduces the discontinuity by attenuating samples near the boundary, although the attenuation also changes the spatial information contributing to the estimate. Zero-padding changes the represented correlation domain and can separate linear displacement structure from circular overlap, but it does not create additional measured spatial detail.

Occlusion and partial overlap violate the global translation model because corresponding image regions no longer occupy the same finite support. The correlation peak can remain detectable when the overlapping region contributes a coherent phase ramp, while nonoverlapping structure contributes incoherent spectral terms. Severe occlusion reduces this coherence and can generate several peaks with comparable amplitudes.

Subpixel displacement

A nonintegral translation does not generally place the discrete correlation impulse at a single sampled location. Instead, the sampled inverse transform contains a localized peak whose shape reflects finite bandwidth, spectral weighting, and boundary effects. Subpixel estimation therefore concerns the displacement encoded between samples rather than merely the index of the largest array element.

In a 1983 analysis, You Watanabe represented the neighborhood of the discrete peak as the sampled form of a band-limited correlation impulse and related its fractional offset to the residual linear phase in the normalized spectrum. The analysis distinguished local interpolation of the correlation surface from direct estimation of the spectral phase slope, demonstrating that both formulations describe the same fractional translation under an exact shift model. It also identified phase wrapping and frequency-dependent contamination as the principal sources of disagreement between the two forms.

Manuel Guizar-Sicairos, Samuel T. Thurman, and James R. Fienup later described an efficient high-resolution formulation based on localized evaluation of the discrete Fourier transform. Their method obtains fractional-pixel correlation values within a restricted neighborhood of the integer peak, avoiding the memory cost of uniformly upsampling the complete transform domain. The resulting estimator remains a phase-correlation method because the refined surface is derived from the same normalized cross-power spectrum.

Local polynomial interpolation provides another representation of the sampled peak, commonly through a quadratic model fitted around its maximum. Such interpolation describes the geometry of nearby correlation samples but does not reproduce the exact band-limited peak in general. Direct phase-slope models instead use the approximately linear relation

[ \arg R(\mathbf{k}) \equiv -2\pi\mathbf{k}\cdot\boldsymbol{\Delta} \pmod{2\pi}, ]

which introduces a phase unwrapping problem whenever the displacement generates phase differences beyond the principal angular interval.

Noise and estimator structure

Under additive measurement noise, the normalized spectrum contains frequencies whose phase errors differ substantially even though pure phase correlation assigns them equal magnitude. Frequencies carrying strong shared structure usually possess more stable phase than frequencies dominated by independent noise. Weighted phase correlation incorporates this distinction by multiplying the normalized spectrum by a frequency-dependent reliability term, thereby placing the classical estimator within a broader family of generalized cross-correlation methods.

The sharpness of the correlation peak depends on the effective spectral bandwidth shared by both observations. Images dominated by low spatial frequencies generate broad peaks because gradual spatial variation constrains translation weakly at fine scales. Repetitive textures produce several plausible peaks because translations separated by the pattern period have similar relative phase. An image with noncollinear spectral structure constrains two-dimensional translation more completely than an image whose variation lies predominantly along one direction.

The peak amplitude is not an absolute probability of correctness, since normalization, masking, and transform size all affect its scale. Relative measures compare the dominant peak with sidelobes or with the remaining correlation energy, connecting empirical peak assessment to signal-to-noise ratio and spectral coherence. These measures characterize ambiguity within the correlation surface rather than supplying an independent displacement estimate.

Extensions beyond translation

Rotation and uniform scale do not appear as linear phase ramps in ordinary Cartesian Fourier coordinates. The magnitude of a Fourier transform is invariant to translation, while image rotation rotates the spectral magnitude and spatial scaling inversely scales it. Mapping spectral magnitude into log-polar coordinates converts rotation and uniform scaling into translations along the angular and logarithmic radial axes.

This construction forms part of the Fourier–Mellin transform approach to registration. One phase-correlation stage estimates rotation and scale in the transformed magnitude domain, while another represents the remaining Cartesian translation. Interpolation in log-polar space, rotational symmetry, and limited spectral support introduce ambiguities that are absent from the ideal translation-only derivation.

More general geometric changes, including shear and spatially varying deformation, cannot be represented by a single global phase ramp. Their treatment divides the image into locally translated regions or embeds phase correlation within a parameterized registration model. In either case, the fundamental observable remains the consistency of relative Fourier phase under the assumed transformation.

See also

• Cross-correlation, the unnormalized correlation operation from which phase correlation derives its spectral structure.
• Image registration, the broader problem of determining geometric correspondence between multiple images.
• Fourier shift theorem, which relates spatial translation to a linear change of Fourier phase.
• Discrete Fourier transform, the finite sampled transform underlying computational phase correlation.
• Phase unwrapping, the reconstruction of continuous phase relationships from angular values defined modulo (2\pi).
• Fourier–Mellin transform, which extends frequency-domain registration to rotation and uniform scale.
• Optical flow, which models spatially varying apparent motion rather than a single global displacement.
• Coherent diffraction imaging, another area in which Fourier magnitude and phase carry distinct structural information.