False Reject Rate (FRR)

A high FRR translates directly into unnecessary material waste and labor rework costs. When a conforming product is flagged as defective, it is either discarded (scrap) or sent for a time-consuming manual re-inspection.

• Financial Impact: In high-volume manufacturing, a 1% FRR on a line producing millions of units can represent hundreds of thousands of dollars in lost margin annually.
• Throughput Bottleneck: Rework loops create unpredictable queues, disrupting lean manufacturing flow and reducing Overall Equipment Effectiveness (OEE).

FRR and the False Accept Rate (FAR) are intrinsically linked by the system's decision threshold. Adjusting the threshold to make the system more stringent (lowering FAR) inevitably increases FRR, and vice versa.

• The Trade-Off: A system that rejects everything has a 0% FAR but a 100% FRR. A system that accepts everything has a 0% FRR but a 100% FAR.
• Visualizing the Balance: This relationship is plotted on a Detection Error Tradeoff (DET) curve, which helps engineers select the optimal operating point for their specific risk tolerance.

The primary driver of false rejects is the system's inability to handle the natural, acceptable variation within the 'good' class. The model's learned boundary for 'conforming' is too narrow.

• Environmental Factors: Minor, acceptable changes in lighting, part orientation, or background texture can push a good sample outside the learned decision boundary.
• Process Variation: Legitimate differences in raw material finish or color between batches, which do not constitute a defect, can trigger a false reject if the model was trained on a homogenous dataset.

FRR is calculated from the outcomes in a confusion matrix derived from a test set of known conforming samples.

• Formula: FRR = False Negatives / (False Negatives + True Positives)
• Alternative Name: In statistics, this is known as the Type I Error rate (for the specific null hypothesis that the sample is genuine) or, more commonly in classification, the Miss Rate.
• Example: If a system evaluates 1,000 known-good parts and incorrectly flags 50 as defective, the FRR is 50 / 1000 = 5%.

A static, global decision threshold is a common cause of high FRR. Adaptive thresholding techniques dynamically adjust the acceptance criteria based on context.

• Local Normalization: The threshold can vary across different regions of an image to account for uneven illumination, preventing false rejects in naturally darker areas.
• Multi-Factor Scoring: Instead of a single pass/fail gate, systems can fuse multiple quality scores (e.g., dimensional accuracy, surface finish, color) with a weighted model to make a more holistic and forgiving decision.

While related, FRR is not the same as the generic False Positive Rate (FPR). The terminology is domain-specific and depends on which class is defined as 'positive'.

• Biometrics/Inspection Convention: The 'positive' class is a genuine, authorized user or a conforming product. A false reject is therefore a False Negative.
• Medical Testing Analogy: A false reject is analogous to a medical test failing to detect a disease that is actually present (a false negative), whereas a false accept is a false positive (diagnosing a healthy patient).

The Type II error counterpart to FRR, measuring the percentage of defective or non-conforming products incorrectly classified as good. FAR and FRR have an inverse relationship; tightening thresholds to lower FAR inevitably increases FRR. In security systems, FAR is the critical metric for preventing breaches; in manufacturing, it represents the escape rate of defects to customers.

The point on a Receiver Operating Characteristic (ROC) or Detection Error Trade-off (DET) curve where FRR equals FAR. EER provides a single-number summary of system accuracy, with lower values indicating better performance. It is the standard benchmark for comparing biometric and inspection systems, representing the optimal operating threshold where both error types are balanced.

Two metrics that reframe the FRR/FAR trade-off for imbalanced datasets common in manufacturing where defects are rare:

• Precision: Of all parts flagged as defective, what percentage are actually defective? High precision means low false positives (low FRR).
• Recall: Of all actual defects, what percentage were caught? High recall means low false negatives (low FAR). The F1 Score is the harmonic mean of both, useful when a single optimization target is needed.

A graphical plot of FRR vs. FAR across all possible decision thresholds, typically using a logarithmic scale. Unlike the ROC curve which plots true positive rate, the DET curve directly visualizes the trade-off between the two error types. System designers use DET curves to select an operating point that aligns with business costs—for example, minimizing FRR when scrap costs exceed customer return costs.

The business framework that gives FRR its financial meaning. FRR directly contributes to internal failure costs through unnecessary scrap, rework, and wasted production capacity. A 2% FRR on a line producing 1 million units annually means 20,000 good products are destroyed or re-inspected. CoQ analysis balances this against the external failure costs that FAR represents, guiding optimal threshold setting.