The first question almost every new customer asks is some version of: "Can you catch 100% of defects?" The honest answer is that detection rate is only half the equation. The other half — the half that determines whether your yield numbers survive the installation — is false reject rate. These two numbers move in opposite directions, and the threshold you set between them is the most consequential decision in any vision inspection deployment.
We spend more time on this conversation than on any other aspect of commissioning. Not because the math is hard, but because the implications run across quality, operations, finance, and customer relations simultaneously — and most teams have never had to think about them all at once.
The Detection-Rejection Tradeoff Is Not Optional
Every vision inspection system works by scoring each unit against a learned model and comparing that score to a threshold. Units scoring above the threshold get flagged as defects; units below pass. The threshold position controls everything.
Set it too permissive and you miss real defects — the light surface bruise, the partial label peel, the 3mm seal gap that looks fine under most lighting angles. Set it too tight and you start flagging acceptable variation as defects: the small shadow cast by a package seam, the slight color shift at the edge of a tomato that's perfectly marketable, the label edge that photographs at a marginally different angle because of a 0.2mm conveyor vibration.
In a binary world this feels like a calibration problem — just find the right threshold and you're done. In practice, the defect space is continuous and heterogeneous, and the acceptable variation in good product is also continuous. There is no threshold position that catches every real defect while rejecting zero good units. You are always making a trade, and the trade has a cost.
What "Sensitivity" Actually Costs at Scale
At 400 units per minute on an eight-hour shift, you're inspecting 192,000 units. A 1% false reject rate is 1,920 good units in the waste or rework bin. At a product cost of $0.40 per unit, that's $768 per shift — before you account for the labor cost of the reject handling station, the rework throughput loss, and the downstream packaging re-run.
We worked with a mid-size ready-meal processor running a pouch sealing line at roughly 320 packs per minute. Their initial threshold configuration was set conservatively after a retailer complaint about a seal defect that had slipped through a previous visual inspection pass. Understandable instinct. The result was a 3.8% false reject rate — approximately 730 good packs per hour in the reject chute. Over a single 10-hour production run, that's more than 7,000 packs of otherwise shippable product.
The downstream cost wasn't just the product value. The reject chute fed a manual inspection table staffed by two QA operatives who had to open each pack, inspect the seal, and either approve it for rework or confirm the discard. At the false reject rate we observed, the table was permanently overwhelmed. Real defects — the ones that actually needed attention — were getting lost in the noise of good product that the camera had flagged incorrectly.
Why Per-SKU Profiles Matter More Than a Global Threshold
One of the structural mistakes we see in existing vision installations is a single threshold configuration applied across all SKUs on a line. This might work if every product is identical — same geometry, same color range, same packaging. Food lines are never like that.
A line running three different yogurt pot sizes and two label variants will have inherently different "acceptable" variance envelopes for each SKU. The 400g pot has a larger label surface area, which means the same label skew angle produces a larger absolute pixel displacement. The 150g pot runs faster through the vision zone because it's lighter and the conveyor doesn't need to slow for it. The seasonal variant has a metallic finish label that saturates differently under the same LED ring illuminator.
When you apply a single threshold to all of these, you're setting it either for the hardest-to-inspect SKU (and over-rejecting everything else) or for the average (and missing defects on the sensitive SKUs). Neither is correct.
Per-SKU threshold profiles let you calibrate each product type against its own baseline. You're not lowering quality standards — you're defining what "defect" means relative to the product's own acceptable variation envelope. This is the same logic your human QC team applies when they look at different products: they don't judge a bruised tomato by the same visual standard as a shrink-wrapped chicken breast.
How We Approach Initial Threshold Setting
When we commission a new line, we don't guess at a threshold. We run a structured calibration pass that typically takes two to four hours of production time, depending on the number of SKUs and how much defect sample material the plant team has available.
The process: collect 200-500 images of confirmed-good units for each SKU. Collect as many confirmed-defect examples as available — even 30-50 real defects is useful, supplemented with synthetic augmentation. Run both sets through the model and observe the score distribution. The good units will cluster in a score range; the defect units should cluster in a different range. The threshold lives between those clusters.
Where those clusters overlap is your ambiguous zone — the region where a unit's score is genuinely uncertain. Every product type has one, and its width tells you something about how well the model has been trained. A well-trained model on a visually distinctive defect type (e.g., a bright green foreign leaf against pale chicken breast) will have almost no overlap. A model trying to distinguish a level-2 bruise from natural surface variation on ripe stonefruit will have significant overlap.
In the ambiguous zone, threshold placement is genuinely a business decision: how much escape risk can you tolerate versus how much yield impact can you sustain? We surface this trade explicitly and let the plant team make the call with numbers in front of them, not abstractions.
The Role of Line Speed in Image Quality
There's a separate dimension to this tradeoff that often surprises plant engineers: line speed doesn't just affect how many units you inspect per minute — it directly affects the image quality available for each decision.
At 200 units per minute on a standard conveyor, a unit spends roughly 80-100ms in the inspection zone under typical camera placement. At 600 units per minute, that window shrinks to under 30ms. For a GigE Vision camera running at 100fps with a standard industrial lens, you're looking at 2-3 frames per unit at high line speed versus 8-10 frames at lower speed. More frames mean more opportunities to catch the defect in an optimal orientation; fewer frames mean you're often making a call from a partial view.
The practical implication: when a line runs near its upper speed limit, you need either faster cameras (which means higher cost and more light), larger depth-of-field optics (which requires more illumination to compensate for the smaller aperture), or an acceptance that inspection accuracy at peak speed will be modestly lower than at standard speed. None of these is a fatal constraint, but they need to be designed for, not discovered at go-live.
Threshold Drift and Why You Need to Monitor It
Thresholds don't stay calibrated forever. The model was trained on product from a specific production period under specific conditions. As raw material suppliers change, as seasonal variation shifts produce color or texture, as packaging materials change lot characteristics, the gap between your threshold and the actual defect distribution can shift.
We're not saying threshold drift is inevitable in every deployment — in a highly controlled environment with consistent inputs, a well-calibrated threshold can be stable for months. But the monitoring has to be active. The signal to watch is false reject rate trend over time. If false rejects start climbing without a corresponding increase in confirmed defects, the model's score distribution has probably shifted relative to your threshold. That's the time to recalibrate, not to chase the symptom with ad hoc threshold edits.
The other drift trigger is planned: when you change packaging suppliers, when you add a new SKU, when you change line speed for a new product mix. These are threshold recalibration events by definition, and they should be in the change management process, not discovered retrospectively when yield numbers move.
What Good Looks Like
A well-calibrated vision inspection deployment on a food line typically runs with a false reject rate between 0.3% and 1.2%, depending on product type and defect specification. Detection rate for the target defect class (defined with agreed AQL thresholds at commissioning) should be above 97% for surface-visible defects in good lighting conditions. These numbers aren't aspirational — they're achievable with proper calibration on most product types we work with.
The path to those numbers is not a set-and-forget threshold dial. It's a documented calibration process, per-SKU profiles, active false-reject monitoring, and a clear protocol for recalibration when conditions change. The technology can hit the accuracy targets; the process has to support it staying there.
The calibration conversation we have at commissioning isn't just a technical setup step. It's the moment where quality expectations, yield economics, and inspection capability get aligned. Skipping it — or treating a default threshold as "close enough" — is the single most common reason early vision deployments underperform.