False positives are a trust problem more than a statistical problem. When your AI inspection system flags a good part as defective, the reject bin fills with acceptable parts, operators develop the habit of bypassing the system, and the QA team loses confidence in the detection rate numbers before they've had a chance to become meaningful. A 0.8% false positive rate on a 2,000-part shift means 16 good parts in the reject bin every day. That's a plant engineer's daily frustration in concrete numbers.
At the same time, cutting false positives by raising the confidence threshold indiscriminately trades one problem for another. This article covers practical approaches to reducing false positives while keeping defect recall above 98%.
Why false positives happen in production
Model training on a controlled image set is not the same as model inference on a production line. The gap between them is usually one of these: lighting variability (ambient light changes with time of day, weather, or nearby machine operation), part presentation inconsistency (parts land in slightly different positions on the conveyor, changing the viewing angle relative to the camera), and surface texture variation within good parts (natural variation in material lot, stamping die wear, surface treatments).
If your training data was captured under stable lighting with parts presented in a consistent fixture, your model learned good-part characteristics that are narrower than the real production distribution. Every good part that falls outside the training distribution registers as unusual, which the model interprets as a potential defect.
Augmenting training data to cover production variation
The most effective single intervention for false positive reduction is expanding the good-part image set to cover the actual production distribution. This means capturing training images across: three or four lighting conditions (morning, midday, and night shift if ambient varies), multiple material lots, and parts at the expected range of conveyor position variation (not just the target position).
You don't need a large dataset for this. On most stamped metal applications, 400 to 600 good-part images covering this variation is sufficient to train a model that generalizes to production conditions. Below 200 images, variance in production conditions will drive false positives. Above 800, you're typically in diminishing returns unless part geometry is very complex.
Per-defect-class threshold tuning
Using a single confidence threshold for all defect classes is the most common tuning mistake we see. A surface crack might have confidence scores tightly clustered near 0.95 for real defects, making a threshold of 0.90 very conservative without costing recall. A surface void might have a less discriminated score distribution, where the same 0.90 threshold cuts meaningful recall.
After the first two weeks of shadow mode data, plot the confidence score histogram for each defect class independently. Look at the overlap between true positive scores and the scores generated by the good parts. The threshold should sit in the gap between those distributions for each class. If there's no gap, you have a model training problem, not a threshold tuning problem.
Using a review queue instead of direct reject
For defect classes where the score distributions overlap significantly, a three-state output is more appropriate than a binary pass/fail. Parts below a high threshold but above a low threshold go to a review queue rather than direct reject. An operator reviews the flagged parts in the review queue and confirms or overrides. This confirmation data feeds back into model retraining.
This approach reduces false rejects to near zero while maintaining full recall, at the cost of operator review time. On a typical 2,000-part shift with 0.3% defect rate, the review queue at 2% of parts is about 40 parts per shift to review, versus the 6 real defects. That's a workload one operator can handle in 15 to 20 minutes. Over four to six weeks, the review queue data retrains the model and the queue volume drops as the model becomes more discriminative.
Monitoring false positive rate as a maintenance signal
False positive rate is not a set-it-and-forget-it number. It should be tracked as a shift-level metric, and a sustained increase in false positives is a leading indicator of model drift, usually caused by a change in the production environment rather than a change in actual part quality. Common triggers: lighting fixture replacement, camera cleaning (which changes the effective aperture slightly), die replacement or significant die wear, and material supplier change.
If your false positive rate doubles over two weeks and part mix and production rates haven't changed, the environment drifted. The fix is a model update with fresh image capture under the new conditions, not a threshold adjustment. Adjusting the threshold to compensate for model drift is a short-term fix that trains your team to tolerate a degraded system instead of maintaining it.
What good numbers look like in practice
On stamped metal surface defect detection after proper shadow mode calibration and threshold tuning: false positive rate below 1% is achievable and should be the target. False positive rates between 0.3% and 0.8% are common in well-deployed systems. Defect recall above 98% for defects meeting the minimum size specification is achievable at those false positive rates.
The two numbers move together, and there is no universal optimal threshold. The right trade-off depends on the relative cost of a missed defect versus a false reject in your specific process. That cost calculation is a QA engineering decision, not a machine learning decision.