When lighting changes force retraining a vision model

A trained matcher — object matching, defect classification, presence checks — learns the images you showed it and nothing else. Dim the light and you can usually level the image back with controller intensity or exposure time. Move the light, change its colour or add a diffuser, and the gradients the network learned are gone: no camera setting recovers them, and the model must be retrained offline on a PC and reloaded to the device. That is hours of work, plus the risk of a station running an unverified model. This guide draws the line between the two cases, gives the gray-value band that decides it, works a drift example with the arithmetic shown, and sets out the capture and controller habits that keep retraining rare.

Which lighting changes can a model absorb without retraining?

Brightness-only shifts that you can level back to within ±10% of the training set's mean gray value need no retraining: LED aging, dust on the optics, a small intensity trim, an exposure correction. The network keys on local contrast structure — the ratio between a feature and its background — not on absolute pixel levels. In practice, uniform brightness change is among the best-tolerated corruptions for a trained network, while a moved or reshaped light source is among the worst, because it rearranges every shadow and highlight the model memorized.

As a working rule, a matcher holds its accuracy through a uniform brightness swing of about ±20% once you restore the gray level, and degrades progressively beyond that. The practical test costs nothing: if the live image looks like a dimmer or brighter copy of the training images — same shadows, same glare, same edge direction — level it and keep running. If the scene itself looks different, you are in the next section.

Which lighting changes force a retrain?

Any change to where light comes from or what colour it is — angle, diffusion, working distance, polarization, wavelength, a different light unit — breaks the gradients the network learned, and only retraining fixes it. Brightness can be levelled; geometry cannot. The model has no physics to extrapolate with, only the examples it was trained on.

The common triggers, in rough order of frequency:

• Moving the light. A ring light raised from 100 mm to 140 mm cuts illuminance to about (100 ÷ 140)² = 0.51 — half — and swings shadow direction on every edge. For a small source the inverse-square law gives this; treat it as an estimate for a large ring.
• Adding a diffuser, dome or polarizer. A dome front erases the specular highlights a metal-surface matcher may be using as features. The physics behind cross polarization to cut glare fixes your glare problem and removes the model's evidence at the same time, so retrain after the change, not before deploying it.
• Changing colour or wavelength. Swapping white for blue or red light rewrites which surfaces are bright; the guidance on which LED colour gives the best contrast picks the colour, and a retrain locks it in.
• A new part finish or supplier. A matte batch after a glossy batch is a lighting change from the camera's point of view.
• A new camera or lens. A different pixel size, field of view or focus changes the gradient scale itself. Run the lens and lighting selection test plan before swapping optics on a trained station.

The failure signature is distinctive: confidence drops across all parts, misses cluster on the same feature type, and no exposure or gain setting buys it back. Angle choice is its own subject — see lighting angles: surface or silhouette — but once trained, the angle is frozen.

If the new set-up involves UV — 365 to 405 nm ring lights used for fluorescence and glue inspection — shield the station: near-UV is largely invisible, so the eye does not flinch from it the way it does from bright white light.

How much brightness drift is too much?

Hold a reference patch within ±10% of its training-set mean gray value; once controller intensity and exposure together cannot pull it back into that band, retrain rather than keep chasing the drift. The band is the working margin that covers measurement noise and normal LED aging between checks.

Worked example. At commissioning the reference patch reads a mean gray of 128 on the 8-bit image, with the strobe at 70% intensity on the controller and exposure at 500 µs. Six months later the patch reads 102 — dust on the optics and LED aging have removed about 20% of the delivered light (128 × 0.8 ≈ 102). The band is 115 to 141 (128 ± 10%), so 102 is out.

Level it in configuration, with no hardware touched:

• Intensity: 70% × (128 ÷ 102) = 87.8, so set 88%.
• Exposure: 500 µs × (128 ÷ 102) = 627 µs, so set 630 µs or the nearest camera step.

Either route lands the patch back near 128. Run the golden set (below); if it passes, the model never knew. Motion sets the ceiling on the exposure route: at 250 mm/s part speed, 630 µs smears 250 × 0.00063 = 0.16 mm of blur — if your feature tolerance is tighter than that, spend intensity, not exposure.

How do you lock the lighting so retraining stays rare?

Drive the light from a controller at a fixed intensity, disable auto exposure and auto gain in the camera, and shroud the station from ambient light; those three settings prevent most drift before it starts. A constant-current output holds LED output steadier than a plain voltage rail as the LEDs warm, and a strobed pulse of fixed width makes every exposure identical shot to shot.

The light controllers range covers both drive types — browse the controller finder or start the system configurator at the controller step. A 4-channel digital box with strobe support gives you one intensity register per light, set once and recorded:

The LS-CTL-0070 is a digital control box with 4 channels, 200 W, RS-232 and strobe.

For stability the choice is drive type and power headroom. The LS-CTL-0290 is a constant-voltage controller at 60 W over 4 channels; the LS-CTL-0180 is a constant-current unit, 4 channels, 120 W, for 24 V lights.

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

On the camera side, fix everything the matcher can see: exposure, gain, white balance if colour. The exposure and gain balance reasoning is in industrial camera image quality settings, and if parts move, read exposure time versus shutter speed before lengthening exposure to level gray.

Bolt the light down and mark the bracket. A ring light nudged 10 mm during maintenance is the most common self-inflicted retrain on flat-part stations; the lights finder filters by size, angle and diffusion so a replacement matches the original geometry. A diffused 120 mm unit is a typical flat-part choice:

The LS-RNG-0426 is a diffused ring light, 120 mm diameter, 24 V.

The strongest lock is a smart camera with the light built into the housing: the geometry cannot drift because nothing moves relative to the sensor, and the optics ship as one assembly — browse smart cameras in the camera finder. Brightness still ages, so keep the gray-value log; the geometry, the usual retrain trigger, is designed out.

The LS-SMC-0063 is a 1.6 MP smart camera, 60 fps, with a 16 mm lens and a polarized white light source.

How do you retrain and redeploy without breaking production?

Train on a PC against images from the new light, validate on a golden set before the model touches production, then upload the model and its settings file to the device and restart the application. Budget hours, not minutes: on this class of controller app the training tool runs off-device, the on-device app has no training interface, and its configuration lives in a settings file rather than a dialog.

1. Capture under the new light — at least the size of the original training set, spanning the same poses and finishes. Include the old images only if the old condition still occurs in production.
2. Retrain in the PC tool, and version the model file and the settings file together; a model paired with the wrong config is the classic silent failure at redeploy. The mechanics of loading a new model without a long outage are covered in swapping a neural network model on edge vision hardware.
3. Validate on the golden set: every image above your deployment confidence threshold. One borderline image is a warning; two failures in twenty is a retrain with more data.
4. Upload model and settings file, restart the application, and re-run the golden set live on the device before the first production part.
5. Shadow-run one shift if the platform allows parallel execution; otherwise swap in a maintenance window and keep the previous model file for rollback. The wider deployment picture is in planning edge inference deployment.

If several product classes share the station, training across multiple product classes shows how to keep one dataset honest.

What images should you log to make the next retrain fast?

Log raw, unprocessed images — the 8-bit or 16-bit mono the model actually consumes — for every reject plus a timed sample of passes, stamped with light intensity, exposure, gain and the settings-file version. On many controller apps, saving is an explicit per-trigger action rather than a background stream: build the save trigger into the sequence from day one, or you will have nothing to retrain from when the light finally drifts.

Keep three buckets:

• Golden set: 20 to 50 commissioning images, read-only, never trained into the main set — it is the ruler you measure every future model against.
• Reject log: every failed part, in full resolution.
• Drift samples: one pass image per hour is enough to plot the reference patch gray value against weeks of LED aging.

A station that logs this way turns a retrain from an archaeology project into an afternoon. The capture-side habits are the same ones in the machine vision lighting stability test: measure the light before it measures you.

When is retraining the wrong fix?

When the image changed for a cause you can remove — daylight on the station in the afternoon, a loose lens, a strobe firing out of step with exposure, mains flicker — fix the cause first; a retrained model inherits the same instability and you will retrain again in a month. Retraining answers a changed but now stable condition, not an unstable one.

Work the causes in this order:

• Ambient leak. Shroud the station, then re-check at the brightest shift hour. Daylight mixing with the strobe is the top cause of slow drift on stations near doors and windows; diagnosing stray light and low contrast covers the checks, and a knowledgebase piece on stabilizing lighting changes at the trigger goes deeper on trigger-level cures.
• Flicker. Overhead lighting beating against your exposure produces banding that moves run to run — see camera flicker under artificial lighting.
• Strobe timing. A light firing outside the exposure window gives dark or partial images; strobe light on but image black walks the timing fix, and the pulse and duty limits are in LED strobe duty cycle: matching lights and controllers.
• No control loop. If the station cannot hold gray value between checks, close the loop: closed-loop lighting control for repeatable vision images describes driving intensity from image feedback — the brightness leveling of this guide, automated.

If none of these is present and the light condition has genuinely and permanently changed, retrain — that is the geometry case above, and it is legitimate. Lighting is the half of the system that decides this: why machine vision inspection succeeds or fails on lighting, and the model only ever sees what the light hands it.

What should you check before committing the new lighting?

Run this list before the new light, bracket or retrained model goes live; each line is the evidence for a number this guide already committed to.

• Reference patch within ±10% of the training mean gray, with intensity and exposure recorded.
• Golden set of 20 or more images passes at the deployment threshold, on the device, live.
• Auto exposure and auto gain off; intensity registers saved with the settings-file version.
• Light bracket torqued, position marked, working distance written on the station sheet.
• Strobe pulse width and delay recorded; duty cycle inside the light's rating.
• Shroud fitted; gray value unchanged between darkest and brightest shift hour.
• New training images archived with the settings file they were captured under.

Done honestly, this list is what makes the ±10% band and the ±20% tolerance real numbers instead of hope. To set a station up in this order from scratch, start at the light step of the configurator or contact us with the part, the finish and the line speed.