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
- Image Enhancement: Principles and the Resolution BarrierMethods that attempt to produce an image with higher pixel density than the source. Interpolation-based SR fills in values by estimation; learning-based SR use...
- Video Frame Analysis and EnhancementA family of computational methods that reconstruct a higher-resolution image by combining information from multiple frames that sample the scene at slightly di...