False reject rate in AI vision: what it is, how to measure it, and how to reduce it
False reject rate in AI vision: what it is, how to measure it, and how to reduce it
Your AI vision system just flagged another good part as defective. The line stops. A technician inspects it. Nothing wrong. That's a false reject, and if it's happening more than half a percent of the time, your inspection system is costing you more than it's saving.
False reject rate (FRR) is one of two error types that define how well an AI vision system actually performs in production. Understanding both, and knowing which one to optimise, is the difference between a system that adds value and one that creates noise.
FRR vs FPR: the two errors that matter
False reject rate measures how often the system rejects a good part. False pass rate (FPR) measures how often a defective part gets through uninspected.
Both hurt you. High FPR means escapes reach customers. High FRR means unnecessary downtime, wasted throughput, and operators who stop trusting the system.
Most manufacturers focus heavily on reducing FPR, which makes sense. But an FRR above 0.5% is a production problem in its own right. Hypernology's benchmark is below 0.5% FRR while maintaining 99% defect detection accuracy. That target is achievable, but only if you know what's driving your rejects.
How to calculate FRR from your inspection logs
You need three numbers from your logs: total parts inspected, confirmed defective parts, and parts flagged as defective by the system.
The formula is straightforward:
FRR = (False rejects / Total good parts inspected) x 100
A false reject is any part the system rejected that a human inspector later confirmed was within spec. Pull a week of data, cross-reference system flags with re-inspection outcomes, and you have your baseline FRR.
If your system doesn't log re-inspection outcomes, start there. You cannot reduce a number you aren't measuring.
The confidence threshold trade-off
Every AI vision model assigns a confidence score to each inspection decision. The threshold you set determines when a score is high enough to pass or low enough to reject.
Raise the threshold and more parts get rejected, because the model is being more conservative. FRR goes up, FPR goes down. Lower it and the opposite happens.
This trade-off is real, but it's not fixed. A well-trained model on high-quality data produces tighter, more reliable confidence scores. That's where the room to improve actually lives.
4 practical methods to reduce FRR without increasing false passes
1. Fix the lighting first
Inconsistent lighting is the single most common cause of high FRR in production environments. Shadows, reflections, and variation between shifts create image noise that the model reads as anomalies. Standardise your lighting setup before adjusting any model parameters. This alone can drop FRR significantly without touching the model.
2. Augment your training data
If your model was trained on images from a single shift, a single camera angle, or a narrow range of acceptable surface variation, it will over-flag parts that fall within spec but look slightly different. Augmenting your training dataset with 1,000 or more representative images per defect class, including diverse examples of good parts, gives the model a more accurate picture of what passing looks like. The result is fewer false rejects and tighter confidence score distributions.
3. Calibrate confidence thresholds to your defect profile
Not all defect types carry the same risk. A scratch on a cosmetic surface is not the same as a dimensional error on a structural component. Set threshold bands by defect category rather than applying a single threshold across the board. This lets you stay strict where it matters and more permissive where the risk is lower, without changing how the underlying model operates.
4. Run structured retraining cycles
A model deployed six months ago is working with data from six months ago. If your materials, suppliers, or production conditions have shifted, your model may be rejecting parts it would have passed earlier because the distribution of good parts has changed. Quarterly retraining cycles using current production data keep the model aligned with what your line is actually producing.
What the numbers look like in practice
Teams that address all four areas, lighting standardisation, training data quality, threshold calibration, and retraining cadence, typically see false positive reduction in the range of 60 to 80% compared to rule-based vision systems. FRR targets below 0.5% become consistent, not occasional.
The path from high FRR to a stable, trusted inspection system is operational. It doesn't require new hardware in most cases. It requires the right data, the right setup, and a system that's built to be tuned.
If your current FRR is above 0.5% and you want to understand what's driving it, the 7-day AI vision deployment guide walks through the setup steps that affect accuracy from day one. For broader context on what defect detection actually covers in a manufacturing environment, the complete defect detection guide is a useful reference.
You can also see how Hypernology's approach compares to rule-based vision vendors on the comparison page.
If your inspection logs are showing FRR above 0.5% and you'd like a second set of eyes on the data, talk to the Hypernology team. Bring your numbers and we'll tell you exactly where the problem is.