Why use green as the reference?

In a conventional Bayer pattern, green is sampled more densely than red or blue, making it a useful structural reference in many methods.

Can a color difference be negative?

Yes. B−G is negative whenever the blue signal is below green on the chosen scale.

Is this a complete demosaicking algorithm?

No. It is a representation and reconstruction idea; neighbor selection, filtering, edge handling and other processing still need to be specified.

AI Glossary / Computer vision and multimodal AI / Intermediate

Color difference (R−G, B−G)

Definition

Color difference signals R−G and B−G represent the red and blue channel values relative to the green channel at the same image location. In demosaicking, these differences can be estimated and interpolated to help reconstruct missing color measurements while sharing useful structural information between channels.

What it means

A typical color-filter-array sensor does not measure a complete red, green and blue triplet at every photosite. Software must estimate missing channels. Interpolating each channel independently can misplace edges and produce false colors. Many natural-image regions contain similar spatial structure in several channels, so their differences can vary more smoothly than the individual channel values. A method can therefore reconstruct green, estimate red-minus-green and blue-minus-green where measurements allow, interpolate those signals, and add them back to green. This entry concerns channel differences in image reconstruction. It is distinct from perceptual color-difference measures, such as distances between two colors in a calibrated color space.

How it works

Begin with raw measurements and the correct color-filter layout. Estimate a full-resolution green channel using an appropriate interpolation method. At a location with a measured red value, subtract the estimated green value to obtain an available R−G sample. Do the corresponding calculation for blue locations. Interpolate the difference signals at locations where they are missing, ideally accounting for image edges. Recover red by adding the estimated R−G value to green, and recover blue by adding B−G.
The order and exact filtering vary between algorithms. Channel correlation is an assumption that helps reconstruction, not an identity that guarantees accuracy. Strong color boundaries, noise and saturation can violate the expected smoothness.

Three practical examples

A raw-photo developer uses green-guided difference interpolation to reduce false color around fine detail compared with a basic independent-channel baseline.
A camera engineer compares difference-based demosaicking methods on colored test patterns, checking whether narrow edges retain their actual color.
A computer-vision team investigates why a detector reacts to colored fringes in processed images. Inspecting raw-channel reconstruction helps distinguish scene content from demosaicking artifacts.

Think of it this way

Think of recording each runner's position relative to a pace runner instead of separately predicting every absolute position. If their relative spacing changes slowly, those differences can be easier to estimate. The analogy fails when a runner changes pace abruptly. Likewise, a vivid color edge can make R−G or B−G change sharply, so smooth interpolation can spread the wrong color across it.

Worked example

Suppose green has already been estimated as 100 at a target location. Two nearby available R−G samples are 20 and 24. Under a simple equal-weight interpolation, the estimated difference is 22, giving a reconstructed red value of 122. If nearby B−G samples are −10 and −14, their average is −12, giving blue equal to 88. The reconstructed triplet is therefore (122,100,88).
These numbers use one consistent linear signal scale and illustrate the add-back operation. They do not specify a complete Bayer demosaicking filter or guarantee that those neighbors are appropriate. If a strong color boundary lies between the neighbors, averaging their differences may create an incorrect intermediate color. Negative difference values are ordinary signed quantities, not necessarily invalid pixel measurements.

Illustrative case study

Fictional example: Elm Camera develops an image pipeline for scanning printed packaging. A difference-based method improves the appearance of several neutral edges, but a saturated red label beside a cyan panel still shows color bleeding. The engineers inspect the difference images and find that the smoothness assumption fails at that boundary. They test directional interpolation and include the label in the regression set. The lesson is to evaluate the assumptions on the actual subject matter, not to treat a visually good result on gray patterns as proof of universal color fidelity.

Common misconceptions and mistakes

R−G is not a perceptually uniform measure of how different two colors look. It is also not automatically the same as a standardized chroma component in a video color encoding. Difference interpolation does not recover information without assumptions, and green-channel errors can influence the reconstructed red and blue channels. The differences need not be constant across every edge. Clipping a negative difference to zero would discard valid information. Finally, raw linear signals and gamma-encoded display values are not interchangeable inputs; the intended signal domain must be understood.

Business use

This concept is useful when selecting or evaluating camera pipelines for inspection, photography and machine vision. Compare demosaicking choices using both visual artifacts and downstream task performance. Keep raw reference captures when possible, record exposure and processing settings, and test fine colored patterns as well as natural scenes. A pipeline optimized for attractive photographs may not be ideal for measuring small colored markings. Explain whether the output is intended for display or quantitative analysis, because processing choices and tolerances differ between those uses.

Check your understanding

Estimated green is 80 and estimated R−G is −15. What red value follows?

Answer: Red is 65. Adding a signed difference reconstructs the channel; the negative difference does not imply a negative red value.

Related terms

Demosaicking
Color filter array (CFA)
Bayer pattern
Color space
color filter array
interpolation
channel correlation
false colour

Learn next

Learn the Bayer sampling pattern and basic interpolation first. Then visualize R−G and B−G on neutral and strongly colored edges. Compare independent-channel and difference-based reconstruction before studying edge-adaptive methods.

Sources

MIT Computational Photography, Demosaicking: https://people.csail.mit.edu/fredo/comp-photo-book/03-basic-image-processing-and-isp-12-demosaicking.html Duran and Buades, A Demosaicking Algorithm with Adaptive Inter-Channel Correlation: https://www.ipol.im/pub/art/2015/145/ Primary educational and research excerpts consulted; full implementation and editorial review remain pending. The numerical illustration and fictional case are original.

Previous: Cologne phonetics Next: Color filter array (CFA)