Exposure, focus and triggering: the three settings that break AI inspection on fast lines

In a Tier-1 automotive parts deployment running 11,520 units per day at 99% detection rate, every inspection recipe locks three values before the model runs: shutter speed, focus position, and trigger offset. Those parameters were measured, calculated, and set during commissioning--not left on factory defaults. The model gets credit for 99% accuracy. The acquisition settings made it achievable. On a correctly commissioned line, the model is the last thing to blame for escapes, because the acquisition parameters were validated first. On most lines where escapes are logged and the model is blamed, the investigation ends at the model. It should not. Most "model drift" tickets are exposure tickets wearing a lab coat.

Why acquisition failures look like model failures

Motion blur, focus drift, and trigger offset errors produce images that the model processes as genuine inputs. The model does not know the image is soft. It classifies the pixel data it receives. If that data no longer contains the edge contrast, sharpness, or positional consistency the training images carried, the model's output degrades--and the degradation is attributed to the model, because the model is the last thing that changed.

Three masking patterns are common. First, shift-dependent performance: detection rate is good on day shift, worse on night shift. The investigation assumes lighting variation or temperature effects on the model. Second, post-changeover escapes: the first 30 to 50 units after an SKU switch show elevated miss rates, then performance recovers. The investigation assumes the model needs warm-up images. Third, SKU-specific underperformance: one product variant consistently shows higher escape rates than others at the same defect threshold. The investigation assumes the training set for that variant is thin.

All three patterns can trace to acquisition settings. Shift-dependent performance traces to auto-exposure drift. Post-changeover escapes trace to focus not updating with the recipe. SKU-specific underperformance traces to focus distance or shutter speed not matched to the variant's surface geometry or height.

Setting 1 -- Exposure: shutter speed and line speed

Shutter speed governs motion blur. While the sensor is integrating light, any part movement in the field of view writes blur into the image. The amount of blur depends on two values: shutter time and line speed.

The formula is direct: blur (in micrometres) = shutter time (in microseconds) x line speed (in metres per second).

Working example: a conveyor runs at 0.5 m/s. The camera resolves 10 µm per pixel at the part surface (a typical figure for a 5-megapixel camera covering a 100 mm field of view). Acceptable blur for edge-level defect detection is one pixel or less -- 10 µm. At 0.5 m/s, a 10 µm blur budget allows a maximum shutter time of 20 µs.

Many industrial cameras default to 1,000 µs (1 ms). At 0.5 m/s, that default produces 500 µm of motion blur per frame -- 50 pixels. Every edge in the image is smeared across 50 pixels. An AI model trained on sharp images cannot reliably detect surface defects in those conditions. The camera is not broken. The default shutter time is wrong for this line speed.

At 1 m/s, the same 10 µm blur budget requires a maximum shutter time of 10 µs. Doubling line speed halves the acceptable shutter window. Lines that were re-specced for faster throughput after initial commissioning often hit this problem: the model that worked at the original line speed now produces unexplained escapes at the new speed.

Achieving shutter times of 10--30 µs with adequate image brightness requires strobe lighting, not continuous LED. A strobe fires a brief, high-intensity pulse at the moment of capture. The sensor integrates a short, bright exposure. Continuous LED at a power level compatible with thermal limits cannot provide equivalent brightness at these shutter times -- dropping shutter time by a factor of 50 requires 50x the light intensity at the same aperture, which exceeds continuous LED thermal ratings.

Auto-exposure is a second exposure failure mode. Cameras with auto-exposure enabled adjust shutter time and analogue gain continuously based on measured scene brightness. If ambient light entering the inspection enclosure changes between shifts -- from skylights, doorways, or overhead lighting that dims after hours -- the camera compensates by extending shutter time. The model receives images with higher blur per shift without any model or configuration change. That is where shift-dependent escape patterns originate.

Fix: set shutter time and gain as fixed values in the inspection recipe. Store them per SKU if surface brightness differs between variants. Enclose the inspection zone from ambient light sources where feasible. Verify the locked values at both shift handovers during commissioning, not just at midday.

Setting 2 -- Focus: depth of field and SKU changeover

Fixed focus at commissioning does not mean fixed focus in production. Three mechanisms degrade focus without any operator change.

Thermal expansion changes lens geometry as the camera warms from a cold start to steady-state temperature over a production shift. For lenses with a short depth of field--necessary for micrometer-level resolution--a 10 degree ambient temperature change can shift the focal plane by enough to reduce sharpness at the part surface. Commissioning a system during the first 20 minutes of a cold shift produces a focus setting that softens as the line warms up.

Part height variation across SKUs is the second mechanism. A lens focused for a part at 200 mm working distance is measurably out of focus for a part at 185 mm. On a high-mix line with 8,000 or more product variants, height differences between SKUs are common. If the focus setting is not updated per recipe, every height-variant SKU receives images softened by the focus mismatch. The degree of softening depends on the lens aperture: wide apertures (low f-numbers) produce shallow depth of field, which amplifies the focus error. Narrow apertures reduce it but require more light, which ties back to strobe requirements.

Continuous autofocus is the third mechanism. Cameras with AF enabled seek the sharpest focus on each frame. On a moving conveyor, the AF system sometimes locks on the belt surface behind the part, or on a position between adjacent parts, producing frame-to-frame focus variation within a single batch. The model sees inconsistent sharpness across images of the same SKU. Training images taken with AF enabled introduce that inconsistency into the training distribution, which makes the model appear to handle soft images -- until production conditions push beyond what the training distribution covered.

The fixes are specific. Disable autofocus on all inspection cameras and physically lock the focus ring at the correct working distance per SKU using a thread-locking compound or set screw. Build a first-article focus check into the recipe handshake: the first part after a changeover triggers a sharpness score against a stored reference before production images are accepted. Set focus at steady-state temperature, not at cold start. On lines where heat output varies significantly between SKUs (e.g., parts arrive at different temperatures from upstream processes), check that the locked focus value is set at the representative operating temperature.

A first-article focus check adds approximately 2 seconds to a changeover. On a line running 8 changeovers per shift, that totals 16 seconds. The alternative is soft images on the first 30--50 units after each changeover, with escapes logged as model-coverage gaps.

Setting 3 -- Triggering: encoder-based vs time-based firing

The trigger determines when the camera fires relative to the part's position in the field of view. A correctly timed trigger places the part at the centre of the frame, within the calibrated inspection zone, at the designed working distance. A trigger that fires 30 ms early or late at 1 m/s places the part 30 mm from the designed position -- roughly one-third of a 90 mm inspection field of view. Defects at the leading or trailing edge of the out-of-position part fall outside the inspection zone and are never evaluated by the model. The system logs a pass.

Two triggering modes are common. Time-based triggering fires at a fixed interval after an upstream sensor detects the part. This works on lines where conveyor speed is constant. When speed varies -- at startup, after a jam clear, during a slow-down for upstream feeding -- the trigger interval no longer corresponds to a fixed position, and part position in the frame drifts. Escapes cluster around line restarts and speed changes, not around specific SKUs, which makes the pattern difficult to diagnose without reviewing trigger timestamps alongside escape logs.

Encoder-based triggering fires at a specific encoder count: a distance traveled since the upstream detection, measured by a rotary encoder on the conveyor drive shaft. This is independent of conveyor speed. As long as the encoder is calibrated (encoder counts per millimetre), the trigger fires at a consistent part position regardless of speed changes. On a line with natural speed variation of +/- 5%, encoder-based triggering maintains sub-millimetre position accuracy; time-based triggering at the same speed variation introduces several millimetres of position error per trigger event.

Encoder-based triggering integrates directly with PLC recipe switching. When the PLC signals a changeover, the trigger distance, fire count, and minimum inter-trigger interval update with the recipe. Parts of different lengths require different inter-trigger intervals to prevent a trailing part from entering the field of view before the current frame is processed. Recipe-based inter-trigger intervals handle this without manual recalibration.

The minimum inter-trigger interval should be set to the expected part separation at maximum line speed plus a 5--10 ms margin. This prevents the previous part from occupying the field of view when the trigger fires for the next part, which produces double-exposure images that the model cannot classify reliably.

Symptom, setting, and fix reference

Applying acquisition discipline to multi-SKU production

The cross-industry proof for this approach is the Tier-1 automotive parts deployment cited at the top of this post. The architecture is explicit: every inspection recipe in the system includes four acquisition parameters alongside the AI model -- shutter speed, focus position, trigger distance (encoder-based), and strobe pulse width. The PLC auto-switching that loads a new model in under 2 seconds for each of 8,000+ product variants simultaneously delivers those acquisition parameters. There is no manual acquisition adjustment at changeover.

That architecture is one reason the system sustains 99% detection across the full SKU range, not only the commissioning SKUs. The throughput figure--270 items per hour versus 60 on the previous hardware-locked system--was achieved by eliminating changeover downtime. Acquisition parameters switching with the recipe is part of what eliminated that downtime: there is no manual exposure or focus readjustment between products. When a new product variant is onboarded, acquisition parameters are set and verified before the first training image is captured. Training images taken with incorrect acquisition settings -- soft focus, motion blur, incorrect part position -- produce a model trained on those images. That model reproduces the acquisition errors it learned from and produces escapes at the same rate.

The same principle applies across any AI vision platform. Acquisition settings must be set, stored per recipe, and validated before training begins. Revising acquisition settings after training requires retraining, which costs hours to a day even with a low-data training system. Correct acquisition at the outset is cheaper than retraining after the fact.

Acquisition settings should be validated and documented at the factory acceptance test stage, before the system reaches the production floor. The vision system SAT acceptance test guide covers shutter speed, focus, and trigger timing as specific checklist items in the FAT and SAT sequence.

When to look beyond acquisition settings

Acquisition settings are not always the root cause of detection shortfalls. Three other causes are common.

Defect-class coverage gaps occur when the training set does not include a specific defect type. The model has no basis to detect it. Correct acquisition settings make the image sharp, but a sharp image of an unknown defect class is still undetected. The fix is additional training data, not acquisition tuning.

Insufficient defect-to-background contrast is a second cause. On certain surface types -- shallow porosity on brushed aluminium, fine scratches on polished stainless -- the grey-level difference between the defect and the background may be 2--5 counts. A correctly focused, correctly exposed camera at this contrast level may not produce enough signal for reliable model output. The fix is lighting redesign: angle, wavelength, or intensity changes to increase defect contrast before the camera records the image.

Distribution-edge variants occur when an SKU's surface characteristics fall outside the range the training set covered. On a high-mix line, occasional variants with unusual surface finishes or geometries appear as out-of-distribution inputs. The model's output confidence drops on these variants. Continuous learning -- flagging low-confidence outputs for human review and feeding them back into the next training cycle -- closes this gap over time. HyperQ AI Vision includes this loop by design; its patented 1,000-image training requirement (versus the 10,000 images required by conventional deep-learning platforms) means adding a new training sample batch for a distribution-edge variant takes hours, not days. That low-data training capability is also why acquisition settings matter more, not less: 1,000 images trained on blurred or misfocused captures constrain the model's performance ceiling from the first production shift.

Where to start on an existing line

On a line with unexplained escapes and a recently retrained model, run an acquisition audit before scheduling another retrain. Capture 50 images at operating line speed. Compare them against images taken at zero speed on the same camera. If sharpness degrades at speed, the root cause is shutter time or triggering, not the model. If sharpness is consistent across speeds but varies by shift, check for auto-exposure drift. If sharpness varies by SKU, check focus distance per product recipe.

That audit takes one production shift. A model retrain does not correct acquisition root causes.

Send your current escape log alongside the shutter speed, focus setting, and trigger mode recorded at commissioning. Within 5 business days, you will receive a structured acquisition audit checklist matched to your line speed, pixel resolution, and SKU count--identifying the specific acquisition parameters to measure and correct before any model retrain is scheduled. No contract required until the audit confirms the model is the correct next target.

Start the acquisition audit