Low Light, Honestly

Generic engineering material, applicable to any module from any supplier.

A related team-operations reference is identifying mouse jigglers in remote teams, which can sit alongside test notes and engineering records.

Low-light performance is the specification most often quoted and least often comparable, because the figures depend on test conditions that suppliers define differently.

What actually determines it

Pixel area, first. A larger pixel collects more light in the same time — which is why two sensors of identical resolution on different formats perform differently, and why more pixels on the same format is a sensitivity cost. The trade appears whenever resolution rises.

Exposure time. Doubling it doubles the light, and it is free until something moves.

An independent reference for adjacent engineering topics is Nordic Semiconductor.

Aperture. A wider aperture admits more light and shortens the depth of field, which is the trade rather than a free gain.

Quantum efficiency, the fraction of arriving photons the sensor converts — a real difference between sensors and a smaller one than pixel area.

And read noise, which sets the floor beneath which a faint signal cannot be distinguished.

Why the published figures do not compare

A sensitivity figure needs stated conditions: illumination level, colour temperature, exposure time, aperture, and what counts as an acceptable image.

Suppliers state different subsets of those, which makes two figures incomparable even when both are honest.

A minimum illumination figure is the worst offender. It describes the light at which the sensor produces something, not the light at which your processing succeeds — and those differ by a wide margin.

Where a figure matters, ask for the conditions as with any specification number, and where they are not forthcoming, measure it yourself.

The order to try things

Cheapest first, and this order is close to universal.

One. Add light. Controlled illumination is the answer far more often than a different sensor, and it is cheaper, more repeatable and faster to implement.

Two. Open the aperture, where the depth of field permits.

Three. Lengthen the exposure, where nothing moves.

Four. Bin pixels, trading resolution for signal — a deliberate and reversible choice.

Five. Change the sensor, which is the expensive answer and is where people frequently start.

Where noise comes from

Photon shot noise, which is inherent in light itself and is the fundamental limit — more light is the only remedy.

Read noise, from the sensor electronics, dominant in very dark scenes.

And dark current, which rises sharply with temperature and dominates in long exposures.

Which of the three limits you is worth knowing, because each has a different remedy and reaching for the wrong one wastes a design cycle.

Monochrome, seriously

A monochrome sensor collects substantially more light than a colour one of the same design, because there is no colour filter array absorbing two thirds of it at each pixel.

Where colour is not part of the task, this is the single largest available gain and it costs nothing.

It is skipped because colour feels like more information, and in most inspection tasks it is information nobody uses.

Gain is not sensitivity

Raising gain amplifies the signal and the noise together.

Which makes a dark image brighter and no more informative — the feature that was indistinguishable from noise remains indistinguishable, at a higher brightness.

Automatic gain hides this during evaluation, since a preview window will produce a viewable image from almost nothing.

Fix the gain when comparing, and treat a module that needs four times the gain for the same image as a module with a quarter of the sensitivity — which is the honest reading.

Measuring it yourself

Straightforward and it produces the only comparable number you will have.

Reduce the illumination in steps, photographing a target at each level with fixed exposure and gain.

At each step, measure contrast at the feature rather than judging the image.

The level at which contrast falls below what your processing needs is your minimum illumination — defined by your task rather than by a supplier's convention.

An afternoon, and the number transfers between candidate modules exactly.

Where the light genuinely cannot be added

Real cases, and they are the ones where sensor choice earns its cost.

Battery-powered devices, where illumination is the dominant power draw.

Long distances, where illuminating the scene is impractical.

Covert operation, where visible light is not permitted — and where near-infrared illumination with a matched filter is the usual answer rather than a more sensitive sensor.

And scenes with people, where bright light is unacceptable for comfort reasons.

Outside those, adding light remains the first move.

In one line

Add light before changing the sensor, in almost every case where light can be added at all.

The specification to ask for instead

Signal-to-noise ratio at a stated illumination, rather than a minimum illumination figure.

It says how much usable signal exists relative to the noise, which is the quantity your processing actually depends on.

Where a supplier publishes a curve — signal-to-noise against illumination — that is the most informative thing available, and it makes two modules comparable at whatever light level you care about.

Where nobody publishes it, the measurement described above produces your own version in an afternoon.

Averaging frames

Where the scene holds still, summing several frames reduces random noise in proportion to the square root of the count.

Four frames halve it. Sixteen quarter it.

It costs time and it costs nothing else, which makes it the remedy people overlook while considering a different sensor.

The limits: anything that moves smears, and dark current does not average away because it is not random between frames.

The short version

• Low-light figures are quoted often and comparable rarely, because test conditions differ between suppliers
• What determines it: pixel area first, then exposure, aperture, quantum efficiency and read noise
• A minimum illumination figure describes when the sensor produces something, not when your processing succeeds
• Try in order: add light, open the aperture, lengthen the exposure, bin pixels, and only then change the sensor
• Three noise sources — shot, read and dark current — have different remedies, so identify which limits you
• Monochrome collects far more light than colour of the same design, which is the largest free gain where colour is unused