Camera pipeline order: analog gain, ADC, then digital gamma

You set the ADC depth to 12 bits, choose a 16-bit pixel format, and the brightest pixel stops at 65,520 instead of 65,535. Raising gamma changes nothing, and the black level setting sits at 0. All three observations come from one fact: a camera's signal pipeline has a fixed order, and each control acts at a different stage. Black level and analog gain work on voltage before the ADC converts it; digital gain and gamma work on counts afterwards. The sections below walk the pipeline stage by stage, show the arithmetic that predicts the saturation value for any bit-depth and container pair, and rank the levers that can still save a blown-out highlight.

The pipeline, stage by stage

Every camera in the area scan cameras range runs light through the same six stages.

1. The photodiode converts photons to charge, and the sensor turns that charge into a small voltage.
2. A black-level offset lifts the voltage slightly above zero.
3. Analog gain multiplies the voltage.
4. The ADC quantizes the voltage into integer counts.
5. Digital processing runs: defect correction, digital gain, gamma or a LUT.
6. The counts are packed into the selected pixel format and shipped over GigE, USB3 or CoaXPress.

Stages 2 and 3 are analog; stages 5 and 6 are digital. That split answers the ordering question outright: gain acts before the ADC, gamma after it, and nothing downstream of a clip can undo it. Camera software mirrors the split. An ADC depth selector and a pixel format selector (they appear as AdcBitDepth and PixelFormat in many drivers) control stages 4 and 6 independently, which is exactly how 12-bit conversions end up inside 16-bit files.

Analog gain: before the ADC

Analog gain is a programmable amplifier between the sensor output and the ADC input, and it multiplies signal and noise together before quantization. Moderate analog gain genuinely helps a dim scene because it uses more of the converter's input range. Excessive analog gain drives highlights past full scale, and the converter pins them at the top code.

Software quotes gain in decibels, and the conversion is worth memorizing:

+6 dB is 1.995×, +12 dB is 3.98×, +20 dB is exactly 10×. Every extra 6 dB roughly doubles the signal and halves the headroom left before clipping.

Worked check: a pixel sits at 40 percent of full scale and you set 12 dB. The amplified level is 0.40 × 3.98 = 1.59 full scales, so the ADC writes its top code and the detail is gone. The same pixel survives +6 dB, where 0.40 × 1.995 = 0.80 stays inside range.

When a camera runs out of analog range it continues with digital gain, which multiplies the counts instead. That brightens the image but adds nothing the ADC never measured, and it consumes container headroom just the same. A gain setting that seems dead is usually auto gain fighting you; the fix is in camera gain settings not applied: turn off auto gain first, and the raw numbers are decoded in camera exposure and gain units: interpreting raw values.

The ADC and black level: where clipping becomes permanent

An ADC with N bits produces 2^N codes: 4,096 codes at 12 bits, numbered 0 through 4,095. Voltage at or above full scale lands on the top code, voltage below the bottom rail lands on 0, and both ends are hard clamps. Once the converter has written 4,095, the information about how much brighter that pixel really was no longer exists anywhere in the camera.

Black level exists because sensor noise wanders below the dark floor. A small positive offset lifts the floor so noise has room to fluctuate without pinning to code 0. Without it, dark pixels stack up at exactly zero and shadows smear together.

A reported black level of 0 normally means you have added no extra offset on top of the sensor's own internal one — not that the floor sits at zero. Cap the lens, take a dark frame, and read the histogram: a floor a few counts above 0 is the offset doing its job, while a hard spike at exactly 0 means dark noise is clipping and shadow detail is being destroyed from the bottom.

Why 12-bit data tops out at 65,520

A 16-bit pixel format is a container with 65,536 slots, numbered 0 to 65,535. It is not a promise of 16-bit precision. When a 12-bit conversion arrives in that container, most cameras shift the value left by four bits — multiply by 16 — so the converter's top code lands at 4,095 × 16 = 65,520.

The four vacated low bits are working room for the camera's own 16-bit arithmetic: digital gain, black-level adjustment and defect patching need space so their results do not overflow. Two visible consequences follow. Untouched 12-bit data steps in increments of 16, and after on-camera processing the low bits can be non-zero — which is why a saturated pixel sometimes reads a few counts above 65,520 instead of exactly on it.

The same arithmetic covers the other pairings: 14 bits in 16 gives 16,383 × 4 = 65,532. A few cameras rescale instead of shifting and fill the container all the way to 65,535, so saturate one frame to learn which convention yours uses. The identical shift explains why 12-bit data in a 16-bit V4L2 frame tops out near 65,520, and the reverse case — a 10-bit sensor delivered in 8-bit images — is covered in ADC bit depth vs pixel format.

Worked example: the saturation value for any pair

For 12-bit data in a 16-bit container: 2^12 − 1 = 4,095, shift = 16 − 12 = 4 bits, scale = 2^4 = 16, saturation = 4,095 × 16 = 65,520. A practical clip detector flags anything at 99 percent of that: 0.99 × 65,520 = 64,864.8, so every pixel reading 64,865 or more counts as clipped.

Spot-check the 10-bit row by hand: 2^10 − 1 = 1,023 and 1,023 × 64 = 65,472. Values in an untouched stream move in steps of the scale factor — sixteens for 12-in-16 — which is a quick way to confirm no on-camera processing sits in the path.

Gamma: digital, monotonic, too late

Gamma is a power-law remap of counts:

In the common convention, a gamma below 1 brightens shadows and compresses highlights; above 1 does the reverse. Some cameras document the exponent as its reciprocal, so run a gradient chart through yours once and note which way the histogram bends.

Two properties decide what gamma can and cannot do. It is monotonic, so an input already at the maximum code maps to the maximum output: saturated stays saturated. And it runs after conversion, so there is no measurement left to recover — gamma reshapes tones that exist and never creates headroom. Digital gain normally sits just before it in the processing chain, and its clipping is equally permanent. When a bright source saturates, gamma is therefore the wrong lever; the setting's full behavior is covered in camera gamma correction: what the gamma parameter changes, and what a count physically represents is the subject of camera pixel values vs voltage: what USB output means.

When highlights clip: the levers that work

Clipping is fixed at or before the ADC, so every working fix lives there too.

1. Shorten the exposure. Brightness scales linearly with exposure time on a linear sensor. The pixel that reads 65,520 at 2,000 µs reads 65,520 × 1,500 / 2,000 = 49,140 at 1,500 µs — safely under the 64,865 clip threshold from the worked example.
2. Cut analog gain. Each −6 dB halves the voltage reaching the ADC; −12 dB pulls the 3.98× example back to roughly 1×.
3. Reduce the light. A lower LED current, a smaller aperture, an ND filter or a shorter pulse all move the same dial.
4. Strobe instead of continuous light. Pulsing the light only during the exposure rejects ambient light and freezes motion at low average power. If a highlight still clips under strobe, cut the pulse current or pulse width — exposure time is no longer the knob.
5. Merge exposures. Take a short frame for highlights and a long frame for shadows and combine them; some cameras do this on-sensor, and software merging works with any two frames.
6. Change hardware last. Sensors whose pixels hold more charge before saturating — full-well capacity, characterized per EMVA 1288 and printed by the sensor maker — clip later, and larger pixels generally hold more than small ones.

Step 4 needs a controller that gates the light hard and cleanly. A four-channel digital unit with a strobe feature and RS-232 control, rated 200 W, fits the job; the LS-CTL-0070 is exactly that:

Pulse-to-pulse brightness repeatability matters too: a drifting pulse current drags your calibrated exposure with it. Driving the light from constant current instead of raw voltage pins that down, and the LS-CTL-0180 is a four-channel constant-current controller delivering 120 W for 24 V lights:

The timing relationship between the camera's exposure window and the light's pulse is its own subject: see camera strobe output timing: exposure pulse and image timestamp. Where the camera's strobe output cannot drive the controller's input, a buffer relay between camera and strobe is the standard fix.

Bench checks before you trust the numbers

1. Turn off auto exposure and auto gain first; auto modes re-adjust mid-measurement and make every reading meaningless. That failure mode has its own guide: camera image brightness drifts over time: turn every auto mode off.
2. Saturate one frame at the light source and read the histogram maximum: 65,520 means the camera shifts, 65,535 means it rescales.
3. Look at the step size in a raw stream: sixteens confirm untouched 12-in-16 data with nothing processing in between.
4. Take a dark frame: a floor a few counts above zero is a healthy black-level offset, a wall at exactly zero is dark clipping.
5. Keep camera gamma off while you measure linearity, and apply your own curve afterwards, where you control the convention.

Cameras for checking this at a bench

Any mono camera shows the raw counts most directly, because no demosaicing step blends clipped channels into their neighbors. You want enough resolution for a meaningful histogram and an interface your PC already speaks; the camera finder filters by both:

The MV-CS200-10UM is a 20 MP mono camera with a 1" sensor on USB3.0. The LS-CMR-0166 is a 12 MP USB3 camera with a 1/1.7" rolling-shutter sensor. The MV-CU020-90GC is a 2.5 MP color model with a global shutter on GigE — color works too, just remember that Bayer interpolation can smear one clipped channel across its neighbors.

The decisions

• A bright pixel pinned at 65,520 (12-bit), 65,532 (14-bit) or 65,535 (rescaled) is expected container arithmetic, not a fault. Cut exposure, gain or light, or merge two exposures.
• Gamma requested as a fix for clipping is the wrong lever; use it afterwards, for tone shaping.
• Black level reading 0 with a dark floor slightly above zero is healthy; change nothing.
• A dim, noisy scene wants more light, a longer exposure or moderate analog gain — then let gamma handle presentation in software.

To spec the camera side, start at the camera configurator; on the light side, the light controller finder and light controller configurator cover gating and constant-current units in the light controllers range. More explainers like this one sit on the cameras topic page, and the applications team answers specific camera questions through the contact page.