On-sensor calibration ===================== The raw output of a pixel cell is not yet ready to be used. A handful of corrections are applied to it before the data leaves the sensor -- partly in the chip's silicon, partly in the driver code that programs the chip -- to deal with imperfections the sensor introduces along the way. They run in a fixed order on every frame: column *fixed-pattern noise* (FPN) trim first, then black-level subtraction, then defect-pixel correction, then lens-shading correction. Knowing what each does matters because the image that reaches user code has already been through all of them. Column-FPN correction --------------------- Each column of the sensor has its own amplifier and column ADC, and small manufacturing variations between them mean every column reads slightly differently from its neighbours. Without correction, this *fixed pattern* shows up as faint vertical streaks in the output -- the streaks stay put from frame to frame because they come from the silicon itself rather than from the scene. The sensor measures the per-column offset and gain trims at the factory, stores them in its calibration ROM, and applies them on every read-out before any further correction runs. Doing this first lets the rest of the pipeline assume every column behaves the same way, including the dark reference pixels that black-level calibration uses next. Black-level calibration ----------------------- The ADC's zero -- the digital count that should correspond to an empty photodiode -- is not perfectly stable. It drifts with temperature, with supply-voltage variation, and slightly from one pixel to the next. Without correction, a perfectly dark frame would not read as zero; each pixel would carry a small positive *dark offset*. The standard fix is to include rows or columns at the edge of the sensor that are physically covered by metal so that no light ever reaches them. Their digital counts give the true dark reference at the current operating conditions. The sensor reads those covered pixels every frame, averages them per row or column, and subtracts the average from every other pixel. The light pixels then come out with zero count for an unilluminated photodiode, regardless of temperature or supply drift. Defect-pixel correction ----------------------- A small fraction of pixels in any sensor are *defective* -- they read a constant value (stuck high or stuck low) regardless of how much light reaches them. Some defects come from manufacturing variation, and more accumulate slowly over the sensor's life (cosmic-ray hits during long periods of operation are the usual culprit). Modern sensors handle this on the fly with a small spatial filter. Every frame, each pixel is compared to its same-colour neighbours; any pixel that lies far enough outside the local median to be implausible is replaced by a value derived from those neighbours. The filter catches both factory defects and ones that develop later, without needing a per-sensor calibrated bad-pixel map, and the defect is invisible in the output. Lens-shading correction ----------------------- The :doc:`cosā“ falloff <../optics/real-lens-effects>` combined with mechanical vignetting from the lens housing gives every uncorrected frame a noticeable corner darkening. The *lens-shading correction* (LSC) hardware on the sensor compensates by multiplying each pixel by a gain that depends on its position in the frame -- 1.0 at the centre, rising smoothly toward the corners to follow the inverse of the measured falloff curve. The sensor provides the multiplier hardware, but the gain map itself is the MCU's responsibility. The driver writes the map into the sensor's LSC registers at start-up, either from a calibration the driver stores or from a fresh measurement against a flat reference target. Some sensors compress the map down to a small set of polynomial coefficients so the on-chip registers can hold it. LSC depends on the lens. Swapping lenses moves the falloff curve, so an LSC map calibrated for one lens will not match another -- a misapplied map looks like dim corners (under-correction) or bright corner blotches (over-correction).