Under a Single Light
Last time I looked at how lens shading flattens the positional unevenness of light, through the frame of the two seats. By correcting the falloff and the color shift at the edges, it brought center and edge under the same light and made the frame even. Only once the frame sits under a single light can we ask: what color is that light, and what do we call white? This installment looks at that question, AWB, through the same frame.
AWB, or auto white balance, is the job of estimating the color of the light illuminating a scene and correcting color so that white looks white. The human eye sees a sheet of white paper under incandescent light and the same sheet in midday shade as equally white. Whether the light is yellow or blue, the brain restores white on its own. A camera sensor has no such correction. It records the light that arrives exactly as it arrives, so white paper under yellow light comes out yellow and white paper under blue light comes out blue. AWB fills that gap, making what people call white come out white on screen.
What AWB Decides
AWB deals with two things in the end. What color the light illuminating the scene is, and what to restore to white under that light.
"What color" is estimation. From the statistics (stats) of the colors that came into the frame, it works backward to the color of the illuminant that lit the scene, usually expressed as a color temperature. Assumptions go into this: that the average of the whole frame is close to neutral (gray world), that the brightest point is white (white patch), or an approach that narrows the illuminant down from the range of colors that sensor could possibly produce. No assumption is always right.
"What to restore to white" is correction. Having estimated the illuminant, you adjust the gains of the R, G, and B channels so that neutral objects look neutral again under that light. But white here is not purely a matter of physics. The white people find natural is often not perfectly neutral but a white that retains some of the light's mood. Evening light reads better with a bit of warmth left in it, and flattening it entirely to neutral looks wrong. So at the end of AWB sits a judgment that sets not "the physically accurate white" but "the white a person accepts as white." Estimation → white point → correction → convergence. That is the skeleton of AWB.
The Core Is Solving a Problem with More Than One Answer
Within that skeleton, what most reveals AWB's character is the "what color": estimating the illuminant. And that work is, by nature, a problem with no single answer.
The color recorded by the sensor is the product of two things: the color an object has (its reflectance), and the color of the light that fell on it. If the frame has a yellow cast, it might be a white object under yellow light, or a yellow object under white light. From a single captured frame you cannot separate the two. AWB has to estimate what the illuminant was on top of information that is, in principle, insufficient. The assumptions mentioned earlier (the average is neutral, the brightest point is white) are devices for filling that gap. But when the scene itself is skewed toward one color, a landscape covered in autumn leaves or a blue sea, those assumptions break and the estimate wavers.
So at the heart of AWB tuning, two judgments are stacked. How to estimate the illuminant from insufficient information, and what to restore to white under the light you estimated. The first is a matter of the estimator's accuracy and stability. The second, where to place the white point between physical neutrality and what feels natural to a person, is closer to a matter of taste.
In the Two Seats, This Work Diverges
The two seats grip the work of estimating the illuminant and setting white from different ends.
In the ISP seat, the asset is the AWB algorithm itself. The estimator that infers the illuminant from insufficient information, the assumptions underneath it and the corrective logic that keeps it from collapsing on color-skewed scenes, the convergence control that carries the white point through changes of light without jumping, and the structure of the calibration interface that accepts a given sensor's color response. These are the handles this seat moves. But this seat does not choose which sensor it will be paired with. The spectral characteristics of the color filter differ from sensor to sensor, and the same light is recorded at different color ratios by different sensors. That color response is a given input, and the estimator must hold up consistently across many sensors' color responses, not one. In this seat, the sensor's spectral characteristics are not a handle you can move but a wall you cannot cross.
In the module seat, that handle and that wall swap places. The AWB algorithm is now a given, arriving with the ISP this module will run on, and the module, already designed and in production with its sensor, color filter, and infrared cut filter, is fixed. What can be moved are the values the algorithm leaves open. You measure the color ratios this sensor actually produces under reference illuminants and seat the white point (calibration); you adjust how the estimate behaves when it wavers on color-skewed scenes to suit this module; and you shift the white point toward the mood of white the client wants, whether to go more neutral or how much warmth to leave in. The sensor color response that one seat could not cross as a wall, the other measures directly under reference illuminants and uses as the starting point of the white point. The same color response is a wall in one seat and a handle in the other.
So "It Went Well" Means Different Things
Both seats share the goal of passing spec. What diverges is what each protects beyond it.
What the ISP seat protects is breadth. An estimator that gets white right on one sensor and one kind of scene isn't enough. It must not collapse badly across many illuminants and many scenes, especially the skewed ones where the assumptions break, and it must offer a clean calibration interface where the integrating side can seat its own sensor's color response. Only then does that estimator get chosen again on the next project. In this seat, AWB "going well" is closer to a generality reused across many inputs.
What the module seat protects is repeatability. This one module passing the client's white balance spec (white accuracy under defined illuminants, no color cast, stability when the light changes) isn't the end. That pass has to hold across the color filter and infrared cut filter variation that shifts from unit to unit in production. Fit the white point too tightly to a single golden sample and the unit on the next line ships with a visible color cast. In this seat, AWB "going well" is closer to a specificity that repeats on this module.
Beyond White
AWB restored white to white. But white being right doesn't mean the rest of the colors are. The red a sensor records is not the red a person sees, and that mismatch doesn't flatten out just by fixing the white point.
I have watched from both sides how the two seats try to protect different things over where white should sit. Next time I move the frame to the whole of color beyond white: the color correction matrix (CCM).
