Is a coded aperture the same as a smaller aperture?

No. Reducing aperture size changes light throughput and depth of field; coding introduces a designed transmission pattern.

Can it estimate depth from one exposure?

Some calibrated coded-aperture systems infer depth-related information from one image, subject to scene and model limitations.

Does it require machine learning?

No. Reconstruction can use classical inverse-problem methods, although learned components may also be incorporated.

AI Glossary / Computer vision and multimodal AI / Advanced

Coded aperture

Definition

A coded aperture is a deliberately patterned opening or mask that changes how light reaches an imaging sensor. In computational photography, a pattern placed in the lens aperture can make defocus blur carry useful information for estimating depth or reconstructing a sharper image with an appropriate computational model.

What it means

A conventional lens aperture is often approximately circular. When a point is out of focus, its recorded blur reflects the aperture's shape as well as other optical effects. Replacing or modifying that opening with a known pattern changes the blur signature. Software can use the known relationship between the pattern, blur and scene to infer information that would otherwise be harder to recover. Coding does not mean encrypting the photograph, and the mask is not a digital filter applied after capture. It alters the physical measurement before the sensor records it. The broader term also includes imaging systems outside ordinary lens-based photography.

How it works

Designers choose an aperture pattern for a specified imaging objective, such as distinguishing different amounts of defocus. They characterize the resulting point-spread function: the sensor image produced by a point-like source. Its scale and shape can vary with depth, focus setting and position in the image. A reconstruction method compares observed measurements with predictions from this calibrated model.
For a locally uniform blur, the observed image can be approximated as the sharp image convolved with a blur kernel, plus noise. Recovering sharp content requires solving an inverse problem, often with assumptions about plausible images. Depth estimation uses the relationship between blur and distance. Real scenes with multiple depths and occluding boundaries require more care than one global blur model.

Three practical examples

A computational-photography research camera uses a patterned aperture to make different defocus levels more distinguishable and support a layered depth estimate.
An inspection prototype evaluates whether known coded blur helps recover useful surface detail away from the focus plane. It must compare that benefit with reduced light throughput.
An imaging laboratory tests alternative mask patterns using calibrated point sources. It measures which patterns retain useful information under the sensor's actual noise conditions.

Think of it this way

Think of recognizing which stamp made a mark by its distinctive pattern. A recognizable coded blur can help software infer how the measurement was formed. The analogy is limited because an image contains many overlapping contributions, not a single isolated stamp. Noise and missing information can make different explanations difficult to distinguish even when the mask is known.

Worked example

Consider a deliberately simplified two-measurement model. A coded system records y₁=x₁+0.5x₂ and y₂=0.5x₁+x₂. If the readings are 8 and 7, solving the equations gives x₁=6 and x₂=4. The two different mixtures preserve enough independent information for this noiseless example.
If both measurements instead equal x₁+x₂, readings of 10 and 10 cannot distinguish 6 and 4 from 3 and 7. This illustrates why the structure of a measurement operator matters. It is not a physical coded-aperture design or a promise of exact photographic recovery. Actual systems involve many pixels, photon noise, calibration error and trade-offs between light collection and distinguishability.

Illustrative case study

Fictional example: Northglass Imaging builds a laboratory camera with a removable patterned aperture. The team photographs a static display containing objects at several known distances and calibrates the blur at each distance. Reconstruction works differently across textured and plain regions, so the team reports uncertainty rather than assigning every pixel a precise depth. It also compares exposures fairly because the mask blocks some light. The experiment supports a design review; it does not establish a finished depth-camera product or prove that coding is useful for every scene.

Common misconceptions and mistakes

A coded aperture does not guarantee an invertible, noise-free image formation process. Even when a mathematical inverse exists, it may amplify noise severely. A pattern that helps depth discrimination may not be optimal for deblurring or light efficiency. Computational reconstruction cannot recover arbitrary detail that the measurement never preserved. Coded aperture and coded shutter are different: the former changes spatial light transmission, while the latter changes exposure over time. Finally, a visually sharper reconstruction is not automatically a more accurate measurement of fine scene structure.

Business use

This concept matters when evaluating specialized cameras and computational imaging proposals. Ask what information the optical code adds, which calibration conditions it assumes and how reconstruction behaves when those conditions change. Compare total capture time, illumination requirements, sensor noise and computational cost with conventional alternatives. Validation should include depth boundaries, low-texture regions and imperfect focus settings. Retain raw measurements and uncertainty information when reconstructed images support measurement or inspection decisions. The investment is in a combined optics-and-software system, so assessing either component alone misses the main trade-offs.

Check your understanding

A mask creates distinctive blur but blocks half the incoming light. Is it automatically better than an ordinary aperture?

Answer: No. The information gained must be weighed against noise, exposure and reconstruction costs under the intended imaging conditions.

Related terms

Computational imaging
Point spread function
Inverse problem
deconvolution
defocus
coded shutter
depth estimation

Learn next

Learn how a point-spread function produces blur, then compare well-conditioned and poorly conditioned inverse problems. Study aperture coding alongside light throughput and sensor noise before interpreting a reconstruction demonstration.

Point spread function

Sources

Levin, Fergus, Durand and Freeman, Image and Depth from a Conventional Camera with a Coded Aperture, project page: https://groups.csail.mit.edu/graphics/CodedAperture/ Primary paper: https://people.csail.mit.edu/billf/papers/CodedAperture-LevinEtAl-SIGGRAPH07.pdf Primary project and paper excerpts consulted; full experimental and editorial review remain pending. The two-equation example is an original conceptual illustration.

Previous: Codec Next: Coding tree unit