Is the PSF the same everywhere in an image?

Only under a shift-invariant approximation or suitable system conditions. Many real systems have spatially varying responses.

Does knowing the PSF remove all blur?

No. Inversion can be unstable, and noise or lost information limits recovery.

Why use a point-like calibration target?

A sufficiently small target helps isolate the system response, but its finite size and measurement conditions still need consideration.

AI Glossary / Computer vision and multimodal AI / Intermediate

Point spread function

Definition

A point spread function, or PSF, describes how an imaging system represents an ideal point source. Instead of appearing as a perfectly sharp point, the source produces a spatial distribution of measured light that characterizes aspects of the system's blur.

What it means

A distant small light can appear as a spot, a diffraction pattern or a streak rather than a single point. The PSF describes that response under specified imaging conditions. Its shape can reflect diffraction, defocus, aberrations, motion and other effects included in the model. A narrower central response can indicate finer localization, but the complete shape, including extended wings, also matters. A PSF is therefore more informative than a single blur-width number. It may change across the image, with wavelength, focus, depth or time.

How it works

For a linear, shift-invariant intensity-imaging model, an extended scene can be treated as many point contributions, each spread by the same PSF. Their sum is expressed as convolution of the ideal scene with the PSF, with noise and other effects added separately. If the response changes with position, one global convolution is only an approximation.
A PSF can be estimated from suitable point-like targets, calculated from an optical model or inferred as part of a reconstruction method. Its normalization and coordinate scale must be specified. Coherent optical systems require care about whether the model concerns complex field amplitude or intensity; a simple intensity convolution should not be applied indiscriminately to every optical setting.

Three practical examples

An astronomy pipeline models how a telescope and observing conditions spread the light from a star.
A microscopy experiment estimates blur from small reference beads under the relevant imaging conditions.
An image-restoration tool uses an assumed or estimated PSF when attempting deconvolution.

Think of it this way

Think of a stamp that turns each ideal point into a small pattern. Stamping and adding the contributions gives a blurred image when the same linear response applies everywhere. The analogy fails when the pattern changes with position, the system saturates or the imaging physics cannot be represented by this simple intensity-addition model.

Worked example

Use a simplified one-dimensional PSF [0.25, 0.5, 0.25], whose weights sum to 1. An isolated point with intensity 100 is represented as [25, 50, 25] across three sample positions. Total intensity remains 100 in this ideal normalized example, assuming the full response is included.
If a second point contributes nearby, its shifted response adds to the first under the linear model. Two close points may therefore merge into one broad-looking feature. Attempting to reverse that spreading is difficult when noise is present or the response suppresses spatial frequencies. Knowing a PSF does not guarantee an exact or stable recovery of every original detail.

Illustrative case study

In a fictional laboratory imaging project, a team measures small bright reference targets near the center and corners of an image. The corner responses are elongated compared with the center response. Applying one center-derived blur model everywhere leaves directional artifacts near the edges. The team investigates a spatially varying model and reviews focus and alignment before interpreting reconstructed detail. This example illustrates the need to match the PSF model to the capture conditions; it does not report performance for a particular instrument.

Common misconceptions and mistakes

A PSF is not always a Gaussian, although a Gaussian can be a useful approximation in some settings. A single full width at half maximum does not describe every feature of the response. The PSF is also not identical to the observed image of any arbitrary bright object; the target's own size and shape can contribute. Another mistake is assuming that a deconvolved image contains only true recovered detail. Noise amplification and model mismatch can produce ringing or other artifacts.

Business use

For quantitative imaging, document how the PSF was obtained, the conditions it represents and whether it varies across the field. Validate restoration on known or independently assessed examples, and inspect faint structures near bright features where response wings can matter. For display-oriented deblurring, compare visual improvement with loss of stability or invented-looking detail. Keep the source image and model settings available so downstream users can distinguish recorded evidence from reconstruction assumptions.

Check your understanding

A normalized PSF spreads an ideal point into several neighboring pixels. Does this necessarily mean that the total signal disappeared?

Answer: No. In an ideal linear, normalized model with the full response included, the signal is redistributed. Clipping, cropping and other real effects can change that conclusion.

Related terms

Convolution
deconvolution
optical transfer function
modulation transfer function
diffraction
defocus
impulse response
spatially varying blur
image restoration

Learn next

Learn convolution and deconvolution next. Compare two PSFs with similar central widths but different tails, then observe how they affect nearby bright and faint features in a simple synthetic scene.

Convolution

Sources

Research example, Point Spread Function Deconvolution Using a Convolutional Autoencoder for Astronomical Applications: https://arxiv.org/abs/2310.19605 Research example, Point spread functions for the Solar Optical Telescope onboard Hinode: https://arxiv.org/abs/0804.4536 Primary research search excerpts consulted; full source and editorial review remain pending.

Previous: Point-set registration Next: Pointer-based model