Detection rate gets all the attention in food vision inspection conversations. What percentage of bruised apples does the system catch? What's the recall on seal defects? These are the right questions to ask, and they're the questions QA managers ask first when evaluating a vision system.
But false reject rate — the percentage of good, shippable units that the system flags as defective and sends to the reject chute — gets discussed far less. This asymmetry is strange, because on most food production lines, the business impact of an excessive false reject rate is large, continuous, and directly measurable.
The False Reject Rate Is a Yield Tax
Every unit that gets falsely rejected is a unit you produced, packaged (in some cases), and then discarded. The cost is the product cost per unit plus any packaging material already applied, plus the downstream labor cost of handling the reject — whether that's a manual re-inspection table, a rework station, or straight to waste.
The numbers add up quickly. At 400 units per minute on a continuous line, a 2% false reject rate is 8 units per minute in the reject chute. Over a 10-hour production run, that's 4,800 good units discarded. At a fully-loaded product cost of $0.35 per unit, that's $1,680 per shift from false rejects alone — before rework labor. Over a production year with two daily shifts and 250 operating days, a 2% false reject rate at this line speed costs roughly $840,000 in product.
This is not a hypothetical edge case. We've seen false reject rates above 3% on new vision installations where the threshold was set conservatively to maximize detection rate, without balancing the yield impact. The plant team typically notices within the first week that the reject chute is running hot, but they often misdiagnose the cause — attributing it to a "real quality problem" in the incoming raw material rather than a threshold calibration issue.
What Drives High False Reject Rates
False rejects originate from three distinct sources, and the fix is different for each.
Threshold set too tight. The most common cause. When a system is calibrated with a detection-rate-maximizing threshold — essentially, when in doubt, reject — it will catch a high fraction of real defects but will also flag borderline good units that happen to score slightly above threshold. The fix is threshold recalibration with explicit attention to the precision-recall tradeoff, accepting a small reduction in detection rate to bring false rejects to an acceptable level.
Training data doesn't represent the full range of good product variation. If the non-defect training examples only show product from a narrow range of conditions — a single production day, a single raw material lot, units from the center of the conveyor where lighting is most even — the model draws a tighter acceptable boundary than the actual product variance justifies. Good product that looks slightly different (different raw material lot, slightly off-center conveyor position, ambient temperature affecting surface texture) scores higher than it should. The fix is expanding the training set for the good product class to capture the full real variation.
Environmental drift that isn't compensated. LED illuminators dim slightly over time. Conveyor surface color changes as belts age. Seasonal variation in raw materials changes product surface properties. Any of these shifts the image distribution away from what the model was trained on. Good product that looked a certain way when the model was trained now looks slightly different and some fraction of it scores above threshold. The fix is either active illumination monitoring with periodic calibration updates, or a scheduled model refresh cycle that retrains on current production images.
Measuring False Reject Rate Accurately
To manage false reject rate, you have to measure it — which means physically checking what's coming off the reject chute, not inferring it from detection rate statistics alone.
The measurement method that gives you clean data: run a 30-minute sampling session at stable production conditions, have a qualified QA inspector independently re-evaluate every unit in the reject bin, and classify each as true defect or false reject. Do this at least once per week during initial deployment and once per month once the system is stable. The ratio of false rejects to total rejects (false reject rate among the rejected population) and the ratio of false rejects to total units produced (false reject rate as a fraction of throughput) are both useful metrics, and they measure different things.
Track the trend over time. A false reject rate that is slowly increasing over several weeks usually indicates model drift due to environmental change or raw material change. A false reject rate that spikes suddenly usually indicates a discrete event — a lighting fixture out, a conveyor belt replaced, a new packaging material lot with slightly different reflectance properties.
Per-SKU Profiles: The Structural Solution
A single inspection threshold applied to multiple SKUs will over-reject some and under-reject others. A yogurt pot that photographs with a slightly reflective lid, run on the same threshold as an opaque cardboard-lidded pot, will generate false rejects from the specular reflection artifacts even though the product is fine. A fresh-cut product with inherently higher visual variation — leafy greens, stonefruit — will generate false rejects if the threshold is calibrated on a less variable product.
Per-SKU inspection profiles let each product type be calibrated against its own baseline and variation range. The threshold for the yogurt pot accounts for the lid reflectance; the threshold for the leafy greens accounts for natural color heterogeneity within an acceptable batch. This isn't lowering the standard — it's defining "defect" correctly relative to what normal looks like for each product.
The operational implication: when a line changeover occurs and a new SKU starts running, the inspection system should automatically switch to that SKU's profile. If the SKU profile doesn't exist yet, the system should alert the operator rather than silently running with an incorrect profile. A vision system running on the wrong SKU profile will generate elevated false rejects or degraded detection, and neither failure mode will be obvious from the line output alone.
The Interaction Between Line Speed and False Reject Rate
Line speed affects false reject rate through image quality. At higher line speeds, the camera has fewer frames per unit and less exposure time per frame. Image quality is marginally lower — slightly higher motion blur, fewer views of each unit. The model makes its decision from noisier data, and the distribution of scores for good units spreads slightly wider. Units near the threshold boundary get pushed above it more often.
In practice, this effect is modest at the line speed ranges most food products run — below about 500 units per minute on a typical conveyor geometry with adequate illumination. Above that, the effect becomes more significant and the optimal threshold shifts. If your line runs at different speeds for different products, the per-SKU profile should encode both the product-specific threshold and any speed-specific compensation needed.
We've measured this directly on a fresh produce sorting line running between 280 and 480 units per minute depending on fruit size. At 280 units per minute, the false reject rate at a fixed threshold was approximately 0.7%. At 480 units per minute with the same threshold, it rose to 1.4%. The detection rate on known defects dropped from 97.2% to 94.8% over the same speed range. These are the numbers that go into the speed-versus-accuracy tradeoff decision at that specific line, with that specific product — and they look different for every deployment.
Reporting False Reject Rate as a KPI
False reject rate should be a first-class KPI on the vision inspection dashboard, alongside detection rate, line speed, and total throughput. Operationally, it belongs on the same screen that the line supervisor watches — not buried in a QA report that gets reviewed weekly.
When false reject rate exceeds a set threshold (a reasonable starting point is 1.5% for most food product categories — adjust based on your product economics), the system should alert the operator. The alert doesn't necessarily mean a line stop; it means someone checks the reject chute and starts the diagnosis process. Most elevated false reject situations have a simple root cause that can be addressed quickly if caught early. Left unmonitored, they accumulate into shift-level yield losses that are much harder to explain after the fact.
Detection rate and false reject rate together define the operating point of your inspection system. Optimizing only one while ignoring the other isn't a quality strategy — it's a calibration error that costs either food safety or yield. Getting both numbers into the acceptable range simultaneously, for each SKU, is the actual goal of vision inspection commissioning.