Skip to content

Super-resolution

Definition

Methods that attempt to produce an image with higher pixel density than the source. Interpolation-based SR fills in values by estimation; learning-based SR uses a trained model to hallucinate plausible detail. The forensic validity of each differs sharply.

Two main approaches
Interpolation-based (estimation) and learning-based (model-generated)
Key limitation
Cannot generate detail not present in any input frame
Works best on
Multi-frame video where objects sample slightly different sub-pixel positions

Common questions

What's the difference between interpolation-based and learning-based super-resolution?+

Interpolation-based methods estimate pixel values by calculating from existing data. Learning-based methods use a trained model to guess plausible details. Their forensic validity differs significantly, so one cannot be treated the same as the other in evidence work.

Can super-resolution create detail that wasn't in the original?+

For multi-frame methods, no. Super-resolution from multiple frames reconstructs higher resolution by combining information from slightly different positions already captured in the input, so it cannot generate detail that was never present anywhere in the original frames. Learning-based single-image methods work differently and can produce plausible-looking detail that has no basis in the original, which is why their forensic validity is treated with much more caution.

Why would forensic examiners use super-resolution on video?+

Video frames often sample a scene at slightly different sub-pixel positions as objects move. Super-resolution combines that information to reconstruct a higher-resolution image, which may reveal details needed for identification or analysis without inventing them.

Related terms

CLAHE
Contrast Limited Adaptive Histogram Equalisation. Divides the image into local tiles, equalises each tile's contrast separately, and blends results. Recovers shadow detail...
Colour calibration
The process of correcting a camera's colour rendering to match a known standard, using a colour reference chart captured under the same...
De-interlacing
The process of reconstructing a full progressive frame from the two interlaced fields captured by analogue video systems. Different algorithms (line doubling,...
Deconvolution
A computational technique that estimates the PSF and inverts its effect to recover a sharper image. The quality depends entirely on how...
Multi-frame averaging
A noise reduction technique that registers multiple frames of the same scene and averages pixel values. Random noise, which differs between frames,...
Nyquist limit
The sampling theorem's constraint: to faithfully record a feature, the sensor must sample it at least twice per cycle. A feature smaller...
Object tracking
Frame-by-frame identification of a target (person, vehicle, or object) across a video sequence, recording its position and trajectory. Used to reconstruct movement...
Point spread function (PSF)
A mathematical description of how the imaging system spreads the light from a single point source across neighbouring pixels. Defocus, diffraction, and...
Spatial resolution
The finest detail a camera system can record, determined by pixel size and lens quality, measured in line pairs per millimetre or...
Temporal synchronisation
The process of aligning the timelines of two or more camera recordings so that events captured from different viewpoints can be placed...

Explained in these topics

Your journey to becoming a forensic professional starts here.

Practice with mock tests, learn from structured notes, and get your questions answered by a global forensic community, all in one place.