Updated July 2026 · 7 min read · Adente Vision Engineering Team
What is a confidence score in AI inspection?
A confidence score is a value, usually between 0 and 1, that an inspection model attaches to each part to express how strongly the evidence supports a decision. For classification it is how sure the model is that a part belongs to a class. For anomaly detection it is often expressed the other way round, as an anomaly score, how far the part deviates from what the model learned as normal. Either way it is a continuous number, not a verdict.
It helps to be clear about what a confidence score is not. It is not a probability that the part is genuinely defective in any calibrated sense, and a score of 0.8 does not mean the part is 80% defective. It is a relative measure the model produces from the patterns it learned, useful for ranking and for setting an operating point, and it only becomes a factory decision once you apply a threshold to it.
Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and every result it produces carries a per-part score that the line turns into an action.
How does a threshold turn a score into a pass or fail?
A decision threshold is the cut-off you set on the score. The rule is simple: a part whose score sits on the pass side of the threshold is accepted, and a part on the other side is rejected or routed for review. That single comparison is what converts a continuous, per-part number into the discrete pass or fail signal a PLC or a robot can act on.
The threshold is a setting, not a fixed property of the model. The same trained model becomes a strict inspection or a lenient one depending only on where you place the line. That is why the operating point should be chosen for the cell, not inherited from a default: a safety-critical part and a cosmetic trim part can share a model but need different cut-offs.
The output stays deterministic once the threshold is fixed. On the line, the accepted-or-rejected result leaves the unit over the protocol the controller already speaks, and a discrete output can drive a reject actuator directly, so the model's judgment reaches the line as a clean pass or fail rather than a raw number the PLC has to interpret.
Why does moving the threshold trade escapes against false rejects?
Moving the threshold trades two errors against each other because they sit on opposite sides of the same cut-off. Loosen it toward accepting more parts and you let more true defects through, which are escapes, or false negatives. Tighten it toward rejecting more parts and you start rejecting good ones, which are false rejects, or false positives. You cannot reduce both at once by moving the threshold alone; you can only shift where the error falls.
This is why the threshold is the single most consequential setting in a deployed inspection. The right position depends on the real cost of each error on your line. Where a missed defect reaches a customer or a safety function, you bias toward catching escapes and accept more false rejects. Where scrapping good product or stopping the line is the expensive outcome, you relax the cut-off and accept a few more escapes. Neither is universally correct, which is the whole reason the score is exposed rather than hidden behind a fixed verdict. For the metric that quantifies the escape side of this trade, see the sibling post on the false-negative rate.
Improving both errors together is a different job: it means better images, better lighting, or a retrained model, not a threshold move.
How does the threshold shape the two error types?
The table shows the direction of the trade, not fixed numbers, because the exact rates depend on your parts, lighting and model.
| Threshold setting | Effect on escapes (false negatives) | Effect on false rejects | When it fits |
|---|---|---|---|
| Loose (accept more) | More escapes get through | Fewer good parts rejected | Cosmetic or low-risk features |
| Balanced | Moderate | Moderate | General quality gating |
| Strict (reject more) | Fewer escapes | More good parts rejected | Safety-critical or customer-critical parts |
| Below the score, routed to review | Escapes reduced | Borderline parts not forced | High-cost mismatch, human check available |
How is per-part confidence used for variant recognition?
Variant recognition is the clearest case where a per-part confidence score earns its place. In a bumper recognition application, the unit classifies each variant with per-part confidence above 0.9, and that score is the number the cell acts on. Above the threshold, the matched program runs. Below it, the part is routed to a safe default rather than forced into a class it is not sure about.
That confidence-first design matters because the cost of a confident wrong answer is high. A borderline part should stop and be checked, not be guessed. Exposing the score per part lets an integrator set the operating point to the cell's real cost of a mismatch, and lets a genuinely ambiguous part fall through to a defined fallback instead of triggering the wrong downstream action.
How does the pass or fail signal reach the line?
Once the threshold turns the score into a decision, the result travels over the protocol your controller already speaks: PROFINET, EtherNet/IP, Modbus TCP, EtherCAT or OPC UA. A pass or fail can drop onto a discrete output for a reject gate, and a variant result can be bit-coded onto outputs or carried as a program-select integer over the fieldbus. The unit also has 4 inputs and 4 outputs at 24V for that discrete signalling.
Because the model, the score and the threshold all run on-device, the decision is made and emitted on the line without a round trip to a server. That keeps the raw image on the line and keeps the pass or fail deterministic and timely for the controller waiting on it.
This post is a spoke of the pillar guide on AI visual inspection; to see the unit that exposes confidence per part on the edge, browse the system overview.