A camera mounted where there happened to be space, under whatever light the plant already had, is not an imaging system — it is a hope that the model will absorb the variance. The pre-pilot site survey exists to remove that hope. It measures the physical imaging conditions on the line before a model is selected, and its output is a per-defect-class record: measured illuminance and its stability across shifts, effective pixels-per-millimetre on the smallest defect you care about, motion blur at rated line speed, and the spread of part pose inside the fixture.
That record is what turns “can vision inspect this?” into a bounded question. Two defect classes with identical appearance in a demo image can land on opposite sides of a go/no-go once you know that one is rendered at four pixels across under a diffuse dome and the other is rendered at twenty under grazing light.
Why does the physical line decide the accuracy ceiling?
Most CV inspection pilots are scoped around the model and the defect list, with the physical line treated as a fixed backdrop. Scoped that way, every source of variance in the imaging chain becomes the model’s problem — and the model can only learn variance it sees in training data. Shift-to-shift lighting drift, a fixture that lets parts rotate eight degrees, a belt that vibrates at the frame rate: each one widens the distribution the model must cover, and none of them appears in the curated set the vendor benchmarked on.
Variance you can engineer out with a lighting bracket is variance you should never ask a model to learn. That is the divergence point between the naive and the expert framing of a site survey. Controllable variance moved into lighting geometry and fixturing is cheap and permanent; the same variance left in the data is expensive, statistical, and re-appears whenever the line changes.
This is the same long-tail failure class the CV inspection feasibility audit is built to catch — benchmark accuracy the line cannot sustain because the line was never characterised. The survey is the physical-evidence stage of that audit. It does not decide feasibility on its own; it supplies the measurements the per-defect-class verdict is written against.
What the survey measures, and in what order
Order matters, because each stage constrains the next. Resolution decides whether the defect is representable at all; lighting decides whether it has contrast; fixturing and conveyor conditions decide whether that contrast survives to the sensor.
| Stage | What you measure | Decision it settles |
|---|---|---|
| 1. Resolution | Effective pixels-per-millimetre on the smallest defect at the working distance, across the full field of view | Whether the defect class is representable at current camera and optics geometry |
| 2. Illumination | Illuminance and colour temperature at the inspection point, sampled across all shifts and ambient conditions | The variance band the model would otherwise have to learn |
| 3. Contrast per defect type | Defect-to-substrate contrast under each candidate lighting geometry (diffuse, grazing, coaxial, backlight) | Which geometry, if any, renders this defect signature |
| 4. Part presentation | Pose spread in the fixture — translation, rotation, tilt — over a representative run | Whether a fixed region of interest holds, or the model must absorb pose variance |
| 5. Motion | Conveyor speed and vibration amplitude against sensor exposure time; resulting blur in pixels | Whether exposure can be short enough at available light, or more light is required |
| 6. Stability | Repeat of stages 2–5 on a second shift and after a changeover | Whether the numbers above are conditions or coincidences |
Every row produces a number, not an impression. In our experience the second-shift repeat is the row teams want to skip and the one that most often changes a verdict — a line that images well at 10:00 on a Tuesday with the roller doors shut is a different optical environment at 03:00 with them open.
Is this defect class limited by the model or by the imaging?
The test is not subtle once the survey numbers exist. If the smallest instance of a defect class occupies fewer pixels than the model architecture can resolve, or if its contrast against the substrate sits inside the sensor’s noise floor under every lighting geometry you tried, the limit is physical. No amount of additional annotated data or a newer detector backbone moves it. If instead a human annotator can reliably mark the defect in the captured frames and the model cannot, the limit is in the model, the training set, or the decision threshold.
Practically, we run this as a two-question check on each defect class:
- Can a trained inspector find the defect in the survey imagery, working only from the frames? If no, the imaging is the blocker.
- Do the survey frames span the lighting, pose and speed variance actually present on the line? If no, any model result from them is a bench result, not a line result.
Specularity is where this most often bites. A scratch on a machined aluminium housing can be invisible under a diffuse dome — the surface returns a broadly uniform bright field — and clearly rendered under low-angle grazing illumination, where the scratch walls throw shadow. Textured substrates behave in the opposite direction: grazing light makes the texture itself high-contrast and buries the defect in it, so diffuse or coaxial illumination is usually the better trade. Low-contrast defects on translucent or thin parts often need a backlight, which changes the fixture, which changes the pose problem. These are engineering choices with a right answer per defect signature, and the survey is where the answer is found rather than assumed.
Speed, vibration and the exposure trade
Motion blur is a straightforward budget and teams still get surprised by it. Blur in pixels is line speed multiplied by exposure time, divided by the millimetres-per-pixel you measured in stage 1. Shortening exposure to hold blur under, say, one pixel reduces the light reaching the sensor, so it must be paid for with brighter illumination, a faster lens, a larger pixel, or strobed light synchronised to part arrival. Vibration adds a second blur term that no exposure setting fixes — it needs mechanical isolation or a trigger point away from the excitation source.
Once you have both numbers, the retrofit question answers itself economically. A strobed LED bar and a machined nest that repeats part pose to within a millimetre are line-side capex with a known price and no ongoing cost. Attempting the same accuracy by training through the variance means collecting defect instances across every lighting and pose combination the line produces — which, for defect classes that occur rarely, can take longer than the payback period of the whole system. On the lines we have surveyed, the lighting-and-fixturing route is usually the cheaper one whenever the variance is mechanically controllable at all (an observed pattern from our engagements, not a benchmarked cost ratio). When the variance is inherent to the product — genuine substrate colour spread across suppliers, for instance — the data route is the only route, and the survey says so.
From survey record to go/no-go
The imaging-condition record hands the feasibility audit three things it cannot get elsewhere: the achievable accuracy band per defect class stated with the conditions under which it holds, the throughput at which that band was measured rather than a bench figure, and a separation of accuracy loss that costs a lighting retrofit from accuracy loss that costs a part or fixture redesign. That separation is the commercially useful output. It is also what makes the audit’s verdict defensible when a defect class is rated no-go — the record shows which physical limit was hit.
The same record has a second life. When the system is hardened, the surveyed imaging conditions become the baseline that inspection reliability evidence is measured against: a drift in illuminance or a fixture that has worn is then a detectable regression rather than an unexplained accuracy dip. We build our computer vision engineering work and the scoping services around that continuity, because a number without its conditions is not evidence.
If you are choosing what to survey first, start with the defect class carrying the largest share of your inspection cost, and measure its imaging conditions before anyone quotes you a model accuracy for it. The uncomfortable case is the one where the survey says the class is unreachable at current camera and conveyor geometry — but knowing that before a pilot is funded is the cheapest version of that news you will ever get.
Frequently Asked Questions
What should a pre-pilot imaging site survey measure, and in what order?
Six stages, in dependency order: effective pixels-per-millimetre on the smallest defect, illuminance and its stability, defect-to-substrate contrast under each candidate lighting geometry, part-pose spread in the fixture, motion blur from conveyor speed and vibration against exposure time, and a repeat of the middle four on a second shift and after a changeover. Resolution comes first because it decides whether the defect is representable at all; the shift repeat comes last because it tells you whether the earlier numbers are conditions or coincidences.
How do you tell whether a defect class is limited by the model or by the imaging conditions?
Ask whether a trained inspector can reliably find the defect in the survey frames alone. If they cannot, the limit is physical — insufficient pixels on the defect, or contrast inside the sensor noise floor under every geometry tried — and no additional data or newer backbone will move it. If they can and the model cannot, the limit sits in the model, the training set, or the decision threshold.
What lighting geometry choices matter for specular, textured and low-contrast defect types?
Specular surfaces such as machined metal often hide scratches under diffuse dome light and reveal them under low-angle grazing illumination, where defect walls cast shadow. Textured substrates behave inversely: grazing light amplifies the texture and buries the defect, so diffuse or coaxial illumination usually wins. Low-contrast defects on thin or translucent parts commonly need backlighting, which changes the fixture and therefore the part-pose problem — the choices are coupled, which is why they are measured rather than assumed.
When is a lighting or fixturing retrofit cheaper than trying to train through the variance?
Whenever the variance is mechanically controllable. A strobed LED bar or a machined nest is one-off capex with a known price and no ongoing cost, while training through the same variance requires defect instances spanning every lighting and pose combination the line produces — which for rare defect classes can take longer than the system’s payback period. When the variance is inherent to the product rather than the line, the data route is the only route.
Does Lighting Fixturing Conveyor Conditions require re-architecture?
Hardware variance trumps algorithmic sophistication in every deployment we’ve profiled: fix the mount before tuning the model.
