Your quality dashboard says 99% detection rate. That number is probably true—and almost certainly misleading.
Detection rate measures performance against the defect categories your system was programmed or trained to find. If your inspection system was configured two years ago, that 99% is measured against a catalogue of defect types that existed two years ago. Every novel defect type introduced since—new material lots, process drift, tooling wear, seasonal variation, new product introductions—exists outside that catalogue.
A system can achieve 99% detection rate against its own catalogue while passing 100% of the defect types it was never designed to catch.
The distinction that matters: your inspection system is detecting 99% of known defects, not 99% of actual defects. The gap between those two figures is where your real quality risk lives. It does not appear on any dashboard. It shows up three months later in your field returns.
Detection rate vs. detection coverage
These are different measurements. Most quality dashboards show the first. Almost none measure the second.
Detection rate (recall): The percentage of true defects correctly flagged—measured against a specific set of defect categories included in the test. High detection rate confirms the system catches what it was designed to catch. It says nothing about what it was not designed to catch.
Detection coverage: The breadth of defect categories the system is capable of detecting—known types, novel types, and variations the process can generate. This is the measurement that predicts whether your field returns will surprise you.
False reject rate: Good parts incorrectly flagged as defective. High false reject rate wastes product and erodes operator trust in the system. Operators who see too many false rejects start overriding the system—which introduces its own escape risk.
Two inspection systems can both report 99% detection rates while having fundamentally different detection coverage profiles. One catches all variations of 15 known defect types. The other catches all variations of 15 known types plus novel anomalies that match no pre-existing category. Same number on the dashboard. Different quality risk in the field.
Why rule-based systems have a built-in blind spot
Rule-based inspection systems encode specific defect templates: a scratch of defined depth, a weld gap above threshold, a color deviation beyond tolerance. These systems are precise within their parameters. They reliably catch what they were told to look for.
The blind spot is structural—not a design flaw, but a consequence of how rule-based detection works. The system can only see what was programmed into it.
Manufacturing reality is not static. Products evolve. Materials shift between suppliers. Environmental conditions change with seasons. Tooling wears incrementally. New SKUs are introduced. Each of these introduces potential defect types outside the existing catalogue.
Rule-based systems are snapshots. They encode the state of defect knowledge at a single point in time. Every change after that—new materials, new suppliers, process drift, tooling wear, seasonal humidity shifts—introduces new failure modes the system was never programmed to recognize.
The longer a rule-based system runs without redesign, the wider the gap between its defect catalogue and the current production reality. The defects that fall through this gap pass inspection cleanly—not because they are subtle, but because no rule exists to catch them.
This is not a criticism of rule-based inspection for stable, well-defined tasks. Where defect types are geometric and repetitive—dimensional checks, presence/absence verification, template matching—rule-based systems work well and are cost-effective. The structural limitation emerges when the production environment changes faster than the inspection catalogue.
Five defect categories and where systems diverge
Not all defect types require the same detection architecture. Understanding which categories your production generates—and which your current system covers—is the first step toward measuring real detection coverage.
1. Surface defects
Scratches, dents, cracks, discoloration, contamination. The most visually apparent category and the one where detection approaches diverge most sharply.
Rule-based handles known, geometric surface variations well. It struggles with novel surface anomalies: a new contamination type from a material lot change, a tool wear pattern that creates surface marks not in the original catalogue, a discoloration specific to a humidity condition that did not exist during system installation.
AI inspection learns what "normal surface" looks like across thousands of samples. Deviations from that baseline are flagged regardless of whether they match a pre-existing category. This is where the detection coverage advantage is largest.
2. Dimensional defects
Parts outside specified tolerances—length, width, diameter, angle. Rule-based systems measure these reliably when geometry is stable and measurement points are defined. AI adds value at tolerance boundaries and for novel geometries where dimensional relationships are complex.
3. Structural defects
Internal voids, porosity, delamination. Detected via X-ray, CT scan, or ultrasonic inspection—not optical vision. Outside the scope of camera-based systems. A complete quality programme includes structural inspection where applicable, but this is not the problem AI vision solves.
4. Assembly defects
Missing components, incorrect orientation, improper fastener placement. Rule-based handles repetitive, fixed-sequence assemblies well. AI handles higher-mix assemblies where the combination space is too large to encode every valid configuration manually.
5. Functional defects
Parts that pass visual inspection but fail under test—electrical continuity, pressure tolerance, mechanical fatigue. Neither rule-based nor AI vision catches these. They require functional test infrastructure. Acknowledge this boundary: visual inspection—any kind—does not replace functional testing.
The categories where AI inspection adds the most incremental value over rule-based are surface defects with irregular or novel patterns, and assembly defects in high-mix environments. Frame this accurately rather than claiming universal superiority.
What changes when you inspect against a baseline
HyperQ AI Vision uses a proprietary segmentation CNN that learns what "normal" looks like for your specific product—at micrometer-level precision, down to 10 micrometres on reflective metal surfaces. The architectural distinction from rule-based inspection:
- Rule-based: "here is what a defect looks like"—encoded in rules at installation
- AI inspection: "here is what normal looks like"—learned from production samples, flags deviations regardless of category
Three capabilities follow from this architecture:
Coverage of unknown defect types
The system builds a baseline from approximately 1,000 production images—where rule-based vision vendors typically require 10,000 to achieve comparable performance on known types. Surface variations that deviate from that baseline are flagged regardless of whether they match any pre-existing defect category.
Novel anomalies. Atypical variations. Defect types introduced by material lot changes or tooling wear. All surface within the detection coverage because the system measures deviation from normal, not match against a catalogue.
Detection plus qualification
HyperQ AI Vision does not just detect—it qualifies. It determines whether a surface variation is acceptable or unacceptable against your specific quality standard. A scratch exists on the surface. Is it within tolerance? Rule-based systems answer: "scratch detected." AI qualification answers: "scratch detected; exceeds allowable depth by 2 micrometres; reject."
This distinction matters operationally. Detection without qualification generates a stream of flags that operators must manually adjudicate. Detection with qualification generates actionable decisions—reducing operator cognitive load and eliminating the subjectivity that makes manual inspection inconsistent.
Adaptation without reprogramming
When process conditions change—new supplier, new material lot, SKU addition—AI inspection adapts through retraining from production samples. With patented low-data training technology, 50 to 200 labelled images add a new product variant after initial deployment. Rule-based systems require reprogramming at each significant change—the same "re-engineering window" that drives changeover cost in high-mix environments.
The practitioner caveat: Lighting is 80% of the job. Any vision inspection—rule-based or AI—requires proper optical setup. Camera selection, lighting angle, background control. The 1,000-image training advantage does not substitute for correct lighting and camera positioning. We select lighting and camera configurations for each customer's specific product geometry and surface properties before any AI training begins.
When to use which approach
The question is not "which system is better." It is: does my current system have the detection coverage my production environment actually requires?
Before evaluating a new inspection system, ask this about your current one: what defect types is my current inspection system capable of catching—and what has changed in my production process since the last time that catalogue was updated?
The decision framework
1. Defect variability: If defect types are geometric, repetitive, and stable—rule-based works well. If defect types are irregular, novel, or variable across material lots and process conditions—AI vision extends detection coverage into the unknown defect space.
2. Product mix and SKU volume: High-mix environments (hundreds to thousands of SKUs) benefit from AI's adaptive retraining versus rule reprogramming at each changeover. A Tier-1 automotive fastener supplier running 8,000+ active SKUs with 12 to 15 product changeovers per shift uses PLC auto-switching to load the correct inspection model in under 2 seconds—zero operator input at changeover.
3. Process stability: Stable, well-controlled processes with infrequent material or supplier changes—rule-based catalogue stays current. Environments with frequent process changes—detection catalogue drifts from production reality over time, creating the escape gap.
4. Consequence of escape: Safety-critical applications (medical devices, aerospace components, automotive braking systems) demand the highest detection coverage—AI vision with human audit at the boundary. The cost of a field escape in these industries is measured in recalls, liability, and regulatory action, not just warranty claims.
5. Line speed: Above approximately 60 parts per minute, manual inspection becomes a bottleneck. Automated vision—rule-based or AI—maintains throughput. HyperQ AI Vision inspects within 0.3 to 1.0 seconds per unit at Full HD resolution.
The three numbers that matter
Every quality review should include these three metrics:
- Detection rate (recall): Percentage of true defects correctly flagged. Always ask: against what defect catalogue? Calibrated when?
- False reject rate: Percentage of good parts incorrectly rejected. Target: below 0.5% for production stability. HyperQ AI Vision achieves 60 to 80% reduction in false positives versus rule-based systems for surface and appearance defects.
- Defect escape rate (PPM): Defective units passing inspection. Best-in-class target: sub-50 PPM. This is the number that predicts field returns—and the number most quality dashboards do not display.
The measurement you are missing
No one can inspect quality into a product. Process improvement prevents defects from being produced. Inspection catches what process improvement has not yet eliminated.
The question is whether your inspection catches a wide enough range of what your process actually produces—including the defect types that were not anticipated when the system was installed.
A 99% detection rate against a two-year-old defect catalogue is a very different number than 99% against your current production reality. The question worth asking before your next quality review: which one is your dashboard actually showing you?
The defect rate you are not measuring shows up in your field returns. Talk to us about measuring it.