AI Glossary / Computer vision and multimodal AI / Advanced

Contrast sensitivity function (CSF)

Definition

A contrast sensitivity function describes how sensitivity to visual contrast varies with properties such as spatial frequency under specified viewing conditions. Sensitivity is commonly expressed as the reciprocal of a detection threshold. A higher sensitivity therefore means that a smaller contrast difference can be detected in the relevant task.

What it means

Two patterns can have the same numerical contrast yet differ in visibility. A pattern's spacing, size and presentation conditions influence whether someone notices it. A CSF helps describe that relationship systematically. In imaging, the concept explains why counting pixels or computing raw differences is not enough to predict every visible change. It concerns a defined visual-detection task; it does not directly tell you whether a person understands a symbol, recognizes an object or prefers an image. Those questions require additional evidence and evaluation.

How it works

A typical experiment varies the contrast of a simple pattern and measures performance under a defined detection procedure. Repeating the procedure across spatial frequencies produces threshold estimates; taking their reciprocals gives sensitivities. Spatial frequency is often expressed in cycles per degree of visual angle. More comprehensive models include factors such as luminance, color, temporal frequency, stimulus area and eccentricity. The cited castleCSF work models several of these dimensions together. Model predictions remain tied to their assumptions and the conditions represented in their supporting data.

Three practical examples

• A display researcher compares how visible small patterned differences are at two viewing distances, recording the angular size rather than only the pixel count.
• A rendering team studies whether a perceptual model can help prioritize image detail, then checks the actual output with representative viewers and tasks.
• A compression researcher evaluates whether numerical errors at different spatial scales have different visibility, keeping the display and viewing conditions controlled.

Think of it this way

Think of a sensitivity chart as a set of differently tuned listening stations, except the signal is visual rather than sound. Some patterns are easier to detect than others even when their physical contrast is similar. The analogy explains selective sensitivity, not a literal set of independent stations inside the eye. It also reminds us that a detection model answers a narrower question than whether the complete picture communicates its intended message.

Worked example

In an original hypothetical experiment, the defined detection threshold is contrast 0.02 for pattern A and 0.05 for pattern B. Their sensitivities are 1/0.02=50 and 1/0.05=20. A contrast of 0.03 is 1.5 times A's threshold but only 0.6 times B's threshold. Under the stipulated model and conditions, it is therefore more favorable for detecting A. This does not mean every viewer will detect A or that a score of 50 is a 50-percent probability. Thresholds summarize performance under a chosen criterion. The numbers are illustrative and do not describe a particular observer, clinical assessment or published CSF dataset.

Illustrative case study

A fictional visualization team notices that fine low-contrast textures in a demonstration look different on a desktop monitor and a headset. Instead of blaming image resolution alone, the team records viewing geometry, luminance settings and the scale of the textures. It uses a CSF model to formulate testable expectations and then runs a task-focused viewing study. When a predicted improvement does not help users identify the intended information, the team revises the design rather than treating the model score as decisive. This scenario illustrates how perceptual modeling can guide experiments without replacing evaluation of the complete user experience.

Common misconceptions and mistakes

Sensitivity and threshold run in opposite directions: a smaller threshold gives a larger sensitivity. Another mistake is treating a single curve as universal across all conditions and observers. Pixels per image are not the same as cycles per visual degree, so viewing geometry matters. A near-threshold detection result also cannot automatically predict the appearance of strong, complex patterns. Be careful when comparing studies that use different detection criteria or stimulus types. A general imaging glossary entry should not be used to interpret an individual's visual health.

Business use

The concept is useful when a team needs to connect display, rendering or compression choices with visibility rather than file size or pixel counts alone. It can help organize an evaluation by making the viewing assumptions explicit. A product decision should still consider the user's actual task, the variety of viewing conditions and the consequences of missed information. Treat the model as one source of design evidence. Keep measured user outcomes alongside model predictions so an optimization does not improve a technical score while making the intended information harder to use.

People also ask

Why take the reciprocal of threshold?

It makes greater sensitivity correspond to detecting smaller contrast changes.

Is a CSF just a screen contrast ratio?

No. It describes visual sensitivity for specified stimuli and conditions, not merely a display's bright-to-dark output ratio.

Does a model apply equally everywhere in the visual field?

Not necessarily; some models explicitly account for eccentricity from the point of fixation.

Check your understanding

Which hypothetical pattern had greater sensitivity?

Answer: Pattern A, because its threshold was lower.

Why is sensitivity 50 not a detection probability?

Answer: It is the reciprocal of a contrast threshold under a defined criterion.

Why record viewing distance?

Answer: It affects the angular scale of the displayed pattern and therefore the relevant spatial frequency.

Related terms

• Spatial frequency
• visual angle
• contrast threshold
• perceptual image quality
• luminance
• foveated rendering

Learn next

Convert a repeated stripe pattern from cycles per image to cycles per visual degree using a specified viewing geometry. Compare threshold and sensitivity plots. Then examine which viewing variables a chosen perceptual model includes before applying its predictions to complex imagery.

More in Computer vision and multimodal AI

• Content-based image retrieval - Content-based image retrieval, or CBIR, finds images using information extracted from their visual content.
• Content-preserving warp - A content-preserving warp is a spatially varying image transformation that balances desired movement against distortion of visible structures.
• Context encoders - Context encoders are neural networks trained to predict missing image regions from the surrounding visible content.
• Contrast-detection autofocus - Contrast-detection autofocus adjusts a camera's focus using image-derived contrast or sharpness measurements.
• Contrast (point operation) - Contrast as a point operation changes the separation between pixel values by applying a tone mapping independently at each image location.
• Contrast stretching - Contrast stretching remaps an image's intensity values so a selected input interval occupies a desired output interval.

Sources

Ashraf and colleagues, castleCSF—A contrast sensitivity function of color, area, spatiotemporal frequency, luminance and eccentricity: https://pmc.ncbi.nlm.nih.gov/articles/PMC10996938/ Primary psychophysical methods study, Spatial four-alternative forced-choice method is the preferred psychophysical method for naïve observers: https://doi.org/10.1167/6.11.13 The thresholds, calculations and visualization-team scenario are original educational illustrations, not clinical guidance or measured study results.

Previous: Contrast (point operation) Next: Contrast set learning