Estimate illuminance from images with luxmeter calibration

A camera can become a useful illuminance estimator for a controlled station, but ordinary image brightness is not a lux measurement. Dark surfaces, automatic exposure and changing light sources can produce different pixel values at the same illuminance. This procedure is for measuring one defined location using a fixed reference target and paired luxmeter readings. You will establish a calibration, test its prediction error and identify conditions that must invalidate the estimate.

Define exactly what the estimate represents

1. Choose one measurement plane and one reference region. State where the luxmeter reading applies, which direction its sensing face points and which target region the camera measures.

Illuminance describes light arriving at a surface, measured in lux. A camera normally records light leaving surfaces toward its lens. Those quantities are related only when surface properties, illumination distribution and viewing geometry are sufficiently controlled.

For an ideal diffuse reflector under fixed illumination conditions:

Here, L is reflected luminance, rho is diffuse reflectance and E is illuminance. Real targets depart from this ideal, so the practical calibration must include the actual target and geometry.

Use a stable, matte reference patch rather than a room-wide average. Fix its position and orientation. The luxmeter sensing surface should represent the same local illumination plane, without the camera or meter casting a shadow on the measurement area.

If the meter occupies the target position during calibration, take sequential readings only after confirming that illumination remains stable. A nearby meter position is acceptable only after checking that the illumination difference is negligible for your error budget.

This procedure estimates illuminance at that reference plane. It does not establish an average room value or illuminance at an unrelated desk, wall or floor position.

Choose acquisition hardware and freeze its response

2. Select an acquisition path that exposes the required settings and image data. Then lock exposure time, gain, aperture, focus and image processing before collecting calibration measurements.

Browse area scan cameras through the camera finder, but confirm feature availability in the selected camera's documentation. Interface type alone does not establish radiometric suitability.

The LS-CMR-0053 is a GigE camera option to assess against your acquisition requirements.

The LS-CMR-0166 provides a USB3 alternative to assess with the same calibration requirements.

Neither example should be treated as a calibrated illuminance meter. Before choosing hardware, confirm whether you can obtain stable image statistics, disable automatic adjustments and retain the settings used for calibration.

Prefer linear image data with documented black-level handling. Disable automatic exposure, automatic gain, adaptive tone mapping and dynamic contrast. For color acquisition, also freeze white balance and channel processing. Read the related guidance on exposure, gain and image processing settings before building the dataset.

A fixed nonlinear response can sometimes support an empirical mapping, but do not assume that a display-ready image is linear. Unknown or scene-dependent processing can make the mapping unreliable.

If considering smart cameras, verify access to a repeatable region statistic and the required settings before ordering. An on-device inspection result is not automatically a radiometric measurement.

Collect paired measurements across the operating range

3. Record image statistics and luxmeter readings under matched conditions. Cover the intended range rather than collecting many nearly identical images.

Use a luxmeter with a suitable measurement range, known calibration status and a documented uncertainty appropriate to the task. The fitted model cannot remove systematic error in its reference instrument.

For each lighting level, record:

• The luxmeter reading and measurement time.
• The image filename and frame timestamp.
• Exposure, gain, aperture and relevant processing settings.
• Target position, camera position and light-source configuration.
• The selected image region and its brightness statistic.
• Any instability, clipping, shadow or scene change.

A practical starting experiment might use 30 distinct illumination conditions, with 10 frames per condition. Reserve 10 conditions for validation and fit the model using the remaining 20. These are planning values, not a guarantee of accuracy.

Keep repeated frames from one condition together when splitting the data. Otherwise, nearly identical frames can appear in both fitting and validation sets, making performance look better than it is.

At each level, allow the light output and camera response to settle. Check the meter before and after the image burst. If the reading drifts beyond the tolerance assigned to acquisition stability, repeat that condition.

For an existing image collection, first audit the metadata and capture procedure. Images with unknown exposure, changing gain or an unrecorded target cannot reliably be combined into this fixed-response calibration. More images do not compensate for missing measurement conditions.

Extract a repeatable statistic from the target

4. Calculate brightness from the same physical patch in every image. Store the region definition with the calibration, not just the final fitted coefficients.

For a stationary setup, a fixed rectangular region may be sufficient. If the patch can shift slightly, locate it using a stable reference and move the measurement region accordingly. Reject images where the target cannot be located confidently.

Keep the region away from target edges, printed markings, shadows and specular highlights. The measurement should describe the reference surface, not a changing mixture of target and background.

For linear monochrome data, an arithmetic mean is a useful starting statistic. Also record the standard deviation and the fraction of clipped pixels. A median can suppress isolated bright defects, but switching between mean and median changes the calibration and requires refitting.

For color data, define the channel calculation explicitly and keep it unchanged. A conventional grayscale conversion is not automatically a photometric response. Its agreement with a luxmeter must still be established experimentally.

Use native image values where practical. Avoid lossy compression and display screenshots. If preprocessing includes resizing or image warping, review antialiasing during downsampling and image warping before assuming that the region statistic is unchanged.

Record dark frames with the lens blocked under the same acquisition settings. These help identify the baseline and its variability. Do not subtract a dark offset during deployment unless the calibration used that same subtraction procedure.

Fit the simplest mapping supported by the data

5. Plot measured illuminance against the chosen image statistic before fitting a model. Look for saturation, a noise-dominated lower end and any change in slope.

Under fixed geometry, spectrum and linear acquisition, a first-order relationship is often a reasonable starting model:

E_hat is estimated illuminance, D is the defined image statistic, a is the fitted slope and c is the fitted intercept. Fit the coefficients using only the training measurements.

Do not force the intercept to zero simply because zero light sounds like zero brightness. Camera black level, dark signal and preprocessing can introduce an offset.

Worked example

Consider illustrative calibration pairs expressed on an 8-bit linear scale. These are arithmetic examples, not product performance measurements.

The slope and intercept are:

For a new mean value of 90:

This mapping applies only over its validated range and under its recorded conditions. Do not extrapolate it toward darkness or clipping.

A polynomial may reduce training error without improving predictions. If a straight-line fit leaves structured residuals, first investigate nonlinear processing, changing source mixtures and saturation. Consider a nonlinear mapping only after identifying why it is needed and testing it independently.

Validate with measurements excluded from fitting

6. Apply the frozen mapping to the held-out conditions. Report error against the luxmeter, not merely how closely the fitted curve follows its training data.

For each validation measurement, calculate:

Relative error becomes unstable near zero illuminance, so report absolute error as well. Useful summaries include mean bias, mean absolute error and the largest observed absolute error within the tested range.

Using the illustrative mapping, suppose a held-out image has a mean value of 94 and its reference reading is 200 lx:

At another condition, a mean value of 166 predicts 390 lx. Against a reference reading of 400 lx, the error is -10 lx, or -2.5%.

Set the acceptance requirement before judging those errors. A process might require an absolute tolerance, a percentage tolerance or a combination. The correct requirement comes from the inspection or measurement task, not from whichever metric looks most favorable.

Check repeatability separately from prediction accuracy. With the example slope, image-statistic variability of 2 units corresponds to approximately 5 lx of estimated variability. That calculation excludes meter uncertainty, target changes and model error.

Repeat validation after a restart and during a separate session. Where supported, verify exposure and gain through frame metadata rather than assuming that saved settings were applied.

Recognize when the mapping stops being valid

A calibration is a conditional measurement model. Its conditions need to accompany every estimate.

A particularly difficult case is a room with independently varying daylight and artificial lighting. Even if total lux remains constant, the camera response can change with the spectral mixture. A single brightness value may no longer identify a unique illuminance.

If the goal is to stabilize inspection images rather than report lux, closed-loop lighting control addresses a different problem. Maintaining a useful image statistic does not, by itself, establish absolute illuminance.

For unrestricted scenes, use a suitable photometric instrument or redesign the measurement around a controlled reference. Do not label an unvalidated whole-image average as a lux reading.

Package the calibration and check before deployment

Store the coefficients together with the image-region definition, preprocessing method, validated range, acquisition settings and calibration date. Keep the held-out measurements and their errors so that later changes can be assessed against the original baseline.

Before accepting a measurement, check that:

• The reference patch is present and correctly located.
• Acquisition settings match the calibration record.
• The region is neither clipped nor below its validated lower limit.
• Target condition and lighting configuration remain acceptable.
• The image statistic falls within the validated interval.
• A recent independent meter check meets the acceptance requirement.

Return an invalid status with a reason when a check fails, rather than a plausible-looking number. Keep acquisition and validity checks separate from the arithmetic that evaluates the fitted mapping.

To specify a station, start with the camera configurator. For help assessing acquisition access and measurement limits, send the target geometry, illuminance range, light-source types and required error tolerance through application support.