How does computational imaging differ from computer vision?

Computational imaging emphasizes forming or reconstructing information from measurements; computer vision emphasizes interpreting visual data. Many systems use both.

Must the raw data look like a photograph?

No. It may consist of coded patterns, projections or multiple measurements that need reconstruction before they become a recognizable image.

Can software replace all high-quality optics?

No. Optics and computation have complementary roles, and the measurements must preserve enough useful information for the intended task.

AI Glossary / Computer vision and multimodal AI / Intermediate

Computational imaging

Definition

Computational imaging combines physical measurement and computation to form images or recover scene information. It treats optics, illumination, sensors and reconstruction algorithms as parts of one system, often using a model of how the measurements were produced to infer the desired result.

What it means

A conventional photograph already involves computation, but computational imaging places reconstruction at the center of the imaging design. The raw measurements may be blurred, coded, indirect or spread across several captures. Software combines those measurements with calibration and assumptions to produce an image, depth map or other representation. The field includes computational photography and broader scientific imaging techniques. It overlaps with image processing and computer vision, yet emphasizes the connection between measurement formation and reconstruction. A visually attractive output is only one possible goal; other systems prioritize measurement accuracy, resolution, acquisition speed or access to information that is not directly visible.

How it works

Designers define the information they want and model how the physical system records it. A forward model predicts measurements from a candidate scene. Reconstruction solves the inverse problem: finding a scene representation consistent with the data and relevant assumptions. Calibration estimates quantities such as optical response, sensor behavior or geometry.
The algorithm may use optimization, analytical inversion, learned models or a combination. In a jointly designed system, the capture process is also chosen to make reconstruction more informative or efficient. Noise, missing measurements and model mismatch limit what can be recovered. Validation must therefore examine both physical acquisition and computation, including cases where the reconstruction prior does not match the scene.

Three practical examples

A burst-photography system aligns several short exposures and combines their measurements to improve a final image under suitable motion and noise conditions.
A coded-aperture prototype modifies the optical measurement so that the recorded blur provides information useful for deblurring or depth estimation.
A microscopy researcher combines measurements taken under different illumination patterns to reconstruct information that a single conventional capture does not directly provide.

Think of it this way

Think of assembling a picture from a set of deliberately chosen clues rather than taking one direct snapshot. The capture system determines the clues, and the reconstruction method interprets them. The analogy is limited because the clues are physical measurements with noise and calibration constraints. An appealing guess is not equivalent to a result supported by those measurements.

Worked example

Suppose a simple sensor records two mixtures of unknown intensities: y₁=x₁+x₂ and y₂=x₁−x₂. If y₁=12 and y₂=4, adding the equations gives 2x₁=16, so x₁=8. Substituting back gives x₂=4. The sensor did not directly report either individual intensity, but the known measurement model permits their recovery.
This is an idealized linear example, not a complete camera design. Real imaging measurements can be noisy, nonlinear or insufficient to determine a unique solution. If both recorded equations were identical, there would be no comparable independent information. The example shows why acquisition design and reconstruction must be considered together.

Illustrative case study

Fictional example: Meadow Optics develops a compact inspection camera whose raw output needs a calibrated reconstruction. The first demonstration looks sharp on a standard target, but an unusual reflective surface creates artifacts. The team compares reconstructed measurements with a reference instrument, studies model mismatch and revises the operating limits. It retains raw data and calibration versions. The case illustrates system-level validation rather than a claim that computation can compensate for every optical limitation.

Common misconceptions and mistakes

Computational imaging is not simply adding a filter to a finished photograph, though the boundaries between fields overlap. It does not necessarily require a neural network. More computation cannot recover arbitrary information absent from the measurements without assumptions. A learned reconstruction can produce plausible features that are not reliable evidence of scene detail. Higher apparent sharpness does not automatically mean higher measurement accuracy. Finally, calibration and illumination are not peripheral setup tasks; they can be central to whether the reconstruction model remains valid.

Business use

Businesses considering an imaging system should evaluate the complete chain from capture to decision. Specify required information, acceptable error, acquisition conditions and failure behavior. Compare performance with conventional alternatives using representative scenes, including unusual materials and motion. Account for calibration maintenance, processing latency, storage and changes to the reconstruction software. If the output supports inspection or measurement, retain traceability from the final result to raw observations and model settings. Procurement based only on a polished demonstration can miss the conditions that make the reconstruction dependable.

Check your understanding

A reconstruction looks detailed but changes substantially when the algorithm's prior changes. What should a measurement user investigate?

Answer: How much of the apparent detail is supported by the observations, how uncertainty is represented and whether the method has been validated against suitable reference data.

Related terms

Learn next

Learn image formation and noise before studying inverse problems. Compare direct measurements with reconstructed quantities, then explore how capture design, calibration and prior assumptions affect what a computational imaging system can reliably recover.

Calibration

Sources

MIT, Computational Imaging course description: https://people.csail.mit.edu/bkph/courses/6870/HWP/6881.html MIT Computational Photography and Video: https://projects.csail.mit.edu/photo/ Primary university teaching and research excerpts consulted; full system-specific and editorial review remain pending. The two-equation example is original.

Previous: Computational graphs Next: Computational intelligence