Visual AI SOP Glossary

A Visual AI SOP project usually involves four departments — process, equipment, quality and IT — and each of them may read the same term differently. The 66 terms below are grouped into eight categories: work standards, vision hardware and imaging, technical implementation, system integration, quality and metrics, project and delivery, algorithms and models, and security and compliance. Each one gives a definition that can be quoted directly, together with the limit of where it applies. Where a term maps to a product capability, the Visual AI SOP System pages show how, and the practical deployment questions are collected in the FAQ.

Work-standard terms

SOP (Standard Operating Procedure)

A written specification of the standard steps, method and requirements for a given task. On the shop floor it usually takes the form of a work instruction, a process sheet or a workstation operating standard. Also known as: Work instruction, Operating standard

Process error prevention

The collective measures that stop a defect being created during an operation, or stop it moving downstream: preventing a missing operation, preventing steps from being performed out of order, and preventing the wrong part being fitted or a part being left out. Also known as: Operation error-proofing, Process-level poka-yoke

Poka-yoke

A concept from the Toyota Production System in which tooling, jigs or system design make a mistake physically impossible, or make it impossible to miss the moment it happens. Visual AI SOP belongs to the second kind: it detects the deviation immediately after it occurs. Also known as: Mistake-proofing, Error-proofing

Missing operation

A product moves on having entirely skipped one of the operations defined in the SOP. Together with out-of-sequence operations and incomplete operations, it is one of the three common categories of process abnormality.

Skipped step

An operator jumps over an intermediate step and performs a later one instead, so the actual sequence no longer matches the standard. It shows up most often when people are chasing output or working from memory. Also known as: Step skipping, Sequence jump

Wrong part and missing part

Fitting a part of the wrong specification (wrong part), or leaving a required part out altogether (missing part). These sit alongside missing operations and skipped steps as the most common assembly exceptions, and they tend to come from look-alike materials being picked in parallel at the same station. Also known as: Misassembly, Missing component

Work-compliance monitoring

Judging whether the operator followed the actions, sequence and protective requirements defined in the SOP — not whether the product is good. It answers "did the person follow the rule", which is a different question from the one product inspection answers. Also known as: Operator compliance monitoring

Standard work element

The smallest action in an SOP that can be observed and judged, such as picking, positioning, fastening or confirming. Breaking a standard down to this level is the precondition for machine vision to check it item by item — a standard that has not been broken down cannot be judged by an algorithm at all. Also known as: Work element

Self-check and cross-check

A self-check is the operator inspecting their own completed operation; a cross-check is the next station inspecting the previous one. Both rely on human diligence, which is why they are the first controls to quietly stop happening. Vision monitoring turns them into automatic records that can be audited after the fact. Also known as: Operator self-check, Cross-check

Human error

A deviation caused by the operator rather than the equipment or the material — a missed step, a reversed sequence, a part fitted the wrong way round. Most SOP failures on a stable line are human error, which is why management measures alone rarely eliminate them. Also known as: Operator error

Vision hardware and imaging terms

Industrial camera

An imaging device built for the shop floor. Compared with consumer cameras it emphasises continuous operation over long periods, external triggering, interchangeable lenses and consistent imaging. Models are split by sensor into area-scan and line-scan, and by shutter into global and rolling. Also known as: Machine vision camera

Global shutter vs rolling shutter

Two ways a sensor exposes. A global shutter exposes every pixel at the same instant, so a moving subject is not distorted; a rolling shutter exposes row by row, which tilts or smears a moving subject. Action recognition on a production line normally uses a global shutter.

Field of view (FOV)

The area the camera covers, set jointly by lens focal length and mounting distance. Too narrow and actions fall outside the frame; too wide and each action occupies too few pixels to judge reliably. It has to be settled during the workstation survey. Also known as: FOV, Coverage area

Lighting design

The choice of light type (ring, bar, backlight, coaxial and so on), mounting angle and intensity. Lighting sets the contrast between subject and background and often does more to determine the floor on recognition rate than the algorithm does. Also known as: Illumination scheme

Motion blur

Smearing caused by the subject moving during exposure. It is eased by shortening the exposure, raising illumination, moving to a global-shutter camera, or timing the capture to the end of a movement rather than the middle of it.

Ingress protection (IP rating)

The rated resistance of an enclosure to solid objects and water. Shop floors generally call for IP54 or better; washdown, oil mist or dust pushes the requirement higher. The rating has to match the actual station, or equipment life drops noticeably. Also known as: IP rating

Spatial resolution (pixel accuracy)

How much of the real scene one pixel represents. What decides whether a feature can be judged is spatial resolution, not the megapixel count — five megapixels spread over a wider field of view resolves less. Size the camera from the smallest feature you must inspect, then pick the sensor. Also known as: Pixel accuracy, Spatial resolution

Depth of field (DOF)

The range in front of and behind the focus plane that stays acceptably sharp. Work is three-dimensional — hands and parts move along the optical axis — so a shallow depth of field blurs the moment the subject leaves the plane. Stations usually stop the lens down to buy depth of field, and pay for it with more light. Also known as: DOF

Technical implementation terms

SOP action recognition

The use of machine vision to recognise the specific actions an operator performs at a workstation — picking, assembling, fastening, inspecting — and from that to decide whether the corresponding operation was carried out. Also known as: Work operation recognition, Workstation action recognition

Edge computing

Running computation and decision-making on devices close to where the data is created, rather than uploading it to the cloud. In a visual SOP context this means recognition happens locally in the plant: no dependency on the public internet, low latency, and production images that never leave the site. Also known as: Edge inference, On-premise inference

Confidence threshold

The minimum confidence a model output must reach before the result is treated as true. Raising the threshold reduces false alarms but increases misses; lowering it does the opposite. It is normally calibrated separately for each workstation. Also known as: Decision threshold, Confidence cutoff

False alarm rate and miss rate

The false alarm rate is the share of normal work wrongly flagged as an exception; the miss rate is the share of real exceptions that go undetected. The two trade off against each other, and the balance is set by how much alarm noise the line will tolerate. Also known as: False positive and false negative, Over-rejection and under-rejection

Cycle time

The standard interval between two consecutive units coming off the line, which reflects the production rate. A visual SOP decision has to complete within the cycle time, which is why inference is usually run locally at the edge. Also known as: Takt time, CT

Workstation calibration

Fixing the camera settings, judgement regions and thresholds for one specific workstation. The same algorithm behaves completely differently under another stations mounting angle and lighting, so calibration is done per station — and the calibration record becomes the reference for reproducing a problem later. Also known as: Site calibration

N-frame confirmation

A decision rule: an action must be recognised across N consecutive frames before it is accepted. It filters out one-off recognition jitter. A higher N suppresses false alarms but adds decision latency, so the value has to stay inside what the cycle time allows. Also known as: Multi-frame confirmation

Allowlist and blocklist

The allowlist holds known non-standard actions that should not raise an alarm; the blocklist holds actions that are explicitly forbidden and alarm immediately. Keeping the two lists separate is how differences that can be ignored are told apart from genuine risk, and it is the main lever for controlling false alarms.

Region of interest (ROI)

The part of the frame that actually takes part in the decision. Restricting the ROI removes interference from irrelevant parts of the image and cuts compute cost, which is what lets one edge device cover more workstations. Also known as: ROI, Judgement region

Trigger (hardware and software)

The signal that tells the camera when to capture. A hardware trigger comes straight from a sensor or a PLC output, so its timing is stable and low-jitter; a software trigger is issued by the application once it decides the state is right, which is more flexible but subject to scheduling. Tight cycle times usually call for hardware triggering. Also known as: Hardware trigger, Software trigger

Image pre-processing

Everything done to an image before the judgement runs: denoising, lens distortion correction, brightness normalisation, ROI cropping. The point is to leave the algorithm only the part that is stable and comparable, so lighting drift and station-to-station differences stop leaking into the result. Also known as: Pre-processing

System integration terms

MES (Manufacturing Execution System)

The information system that manages execution on the shop floor, typically covering work orders, materials, operation routing and quality records. It is the system of record for operation traceability data. Also known as: Shop-floor execution system

Production traceability

The ability to pin a quality problem down to a specific operation by recording when, at which workstation and with what result each step was executed. Visual SOP supplies traceability data at operation level. Also known as: Process traceability, Operation traceability

Process validation

Comparing the operations actually captured against the standard sequence, item by item, to confirm that nothing was missed, no step was skipped and the order is correct. It is the core logic of a Visual AI SOP system. Also known as: Operation verification, Process reconciliation

PLC (Programmable Logic Controller)

The controller at the centre of machine control on a production line. A Visual AI SOP system normally does not write to the PLC and does not take part in interlock control — it collects alongside and only raises prompts. That is the precondition for a plant accepting the system at all. Also known as: Programmable logic controller

OPC UA

A vendor-neutral industrial communication standard widely used to feed equipment data into higher-level systems. Vision results can be published over OPC UA to SCADA or MES, which avoids writing a separate interface for every host system.

MQTT

A lightweight publish/subscribe protocol suited to shop floors with many devices and limited network bandwidth. It is a common transport for reporting vision decisions upward.

Andon system

The shop-floor installation used to call for help and display line status. Vision exceptions can be routed into the existing Andon so the prompt appears where people already look, instead of on one more screen nobody watches. Also known as: Andon

e-SOP (electronic work instruction)

A system that replaces paper work instructions with screens, adding step display, version control and sign-off. It answers whether the standard is being presented correctly; visual SOP answers whether the standard is being followed. The two complement each other rather than replace each other. Also known as: Electronic work instruction

Work order

The smallest unit in which production work is issued, normally carrying the part number, quantity, routing and planned time. Vision results are archived against the work order — that is what makes it possible to answer "how was this batch actually built" later on. Also known as: Production order

Data write-back

Writing station results, timestamps and serial numbers back into the system of record. The fields and the frequency have to be agreed with the MES side in advance: write too much and the cycle slows down, write too little and the traceability chain breaks in the middle. Also known as: Result write-back

Quality and metrics terms

Detection rate (recall)

The share of real exceptions the system actually catches. It has to be read alongside the false alarm rate: detection rate on its own is easy to inflate by loosening the threshold, but the line then drowns in alarms. Also known as: Recall, Catch rate

Precision

The share of raised events that were genuinely exceptions, the complement of the false alarm rate. Low precision means most alarms are noise, and the floor switches the system off before long — the most common way vision projects die.

First article inspection (FAI)

Full confirmation of the first unit after a changeover or a restart. Visual SOP can record how each critical operation was executed on that first article, giving the changeover sign-off something checkable to point at. Also known as: FAI, First-piece approval

First pass yield (FPY)

The share of units that clear every operation without rework or repair. A missed operation often does not show up in FPY straight away, so a stable FPY does not by itself mean the process was followed. Also known as: FPY, First-time-through

PPM (parts per million)

A defect rate expressed in parts per million, commonly used to describe quality at volume. For high-value units such as semiconductors or battery cells, the severity of a single defect matters as much as the PPM figure and is tracked separately. Also known as: Parts per million

Alarm fatigue

What happens once alarm density exceeds what the floor can act on: operators start ignoring alarms, and eventually switch the system off. Capping total alarm volume, splitting by severity, and downgrading non-critical events to records is what keeps a system in use over the long run. Also known as: Alert fatigue

CPK (process capability index)

A measure of how much a process varies relative to the tolerance it has to hold. It has nothing directly to do with vision, but it settles a common argument on the shop floor: whether a fluctuating pass rate comes from the process itself or from work slipping past inspection. Also known as: Process capability index

SPC (statistical process control)

Ongoing statistical monitoring of a process so that a trend is flagged before it becomes scrap. Sampling data is intermittent by nature, whereas vision monitoring produces a judgement on every unit — which can feed SPC a far more continuous signal. Also known as: Statistical process control

Project and delivery terms

Feasibility assessment

The judgement made on a specific workstation before a project starts: whether the action is visible, whether lighting can be controlled, whether the cycle time allows it, and whether the acceptance criterion can be stated clearly. It normally ends in one of three conclusions — doable, not doable, or change something first. Also known as: Workstation assessment

Scenario library

A reusable set of configurations accumulated by industry and by operation, covering judgement logic, calibration parameters and typical exceptions. The thicker the library, the shorter the ramp-up on the next workstation.

Single-station pilot

Deploying at one workstation to prove the result before copying it to the rest of the line. The point of a pilot is not only to validate recognition rates but to find out whether the people on the floor are willing to use it — which is harder than the algorithm and far more often overlooked. Also known as: Pilot workstation

Acceptance criteria

The measurable conditions agreed with the customer before handover, usually covering detection rate, false alarm rate, response time and data completeness. They have to be written down during the pilot, not negotiated at the acceptance meeting. Also known as: Sign-off criteria

SLA (service level agreement)

A written commitment covering response time, repair time, spare parts and availability. Availability of a vision system depends on cameras, lighting and compute hardware, so the SLA has to cover that hardware and not only the software. Also known as: Service level agreement

Proof of concept (POC)

A deliberately small trial run before any purchase decision — usually one or two stations, one product, a bounded period. Its purpose is to answer a single question at the lowest possible cost: does this actually hold up on our line. Also known as: POC, Pilot trial

On-site commissioning and training

The parameter adjustment, threshold calibration and operator training carried out at the customer site during rollout. Vision judgement depends heavily on local lighting, station geometry and part variation, so it does not stabilise when tuned remotely. This phase is the one most often underestimated in delivery schedules. Also known as: Commissioning

Algorithm and model terms

Classical vision vs deep learning

Classical vision judges from hand-designed features — edges, templates, colour — and is explainable with modest sample requirements. Deep learning learns its own features from data, coping with more variation but needing labelled samples. SOP projects generally mix the two: classical methods for locating and presence checks, deep learning for judging whether an action was carried out properly. Also known as: Classical vision, Deep learning

Template matching

Comparing a stored reference image against the current one to locate a target or decide whether something is present. It suits stable shapes under controlled lighting, and remains the most common classical method for station location and presence checks. Also known as: Pattern matching

Data labelling

Marking up image samples by hand before training, telling the model which frames show the correct action and which show a missing part. Label quality sets the ceiling on what the model can achieve, and inconsistent labelling is the most common reason these projects fail — two labellers disagreeing on what counts as "seated properly" leaves the model nothing stable to learn. Also known as: Annotation

Model generalisation

How well a model keeps its accuracy on products, lighting or stations it has not seen. The symptom of poor generalisation is a system that is accurate on the pilot station and fires false alarms the moment it moves next door — which is precisely why a successful single-station pilot is not evidence that the whole line is ready. Also known as: Generalisation

Training set and validation set

The samples used to train a model, and the separate samples held back — unseen during training — used to measure how well it works. The two must stay apart: measuring against samples the model has already seen produces an accuracy figure that is flattering and will not survive go-live. Also known as: Validation set

Training and inference

Training is how a model derives its decision rules from samples, normally offline. Inference is a trained model producing a result from a live image. What runs on the line is inference, and its demands on latency and stability are far stricter than training. Also known as: Inference

Security and compliance terms

Data anonymisation

Masking faces, badges, screen contents and other personal information in video or images before they are stored or transmitted. Shop-floor video contains people by default, so whether and how far to anonymise is a decision for the design stage rather than something to retrofit after go-live. Also known as: Anonymisation

Worker privacy

Station monitoring captures employees at work, which falls under personal data. A defensible approach states the purpose and scope of collection, limits the retention period, restricts who can view the footage, and tells employees in advance. These arrangements normally need HR in the room, not just engineering. Also known as: Employee privacy

Audit trail

A record of who changed a judgement threshold, deleted a record or exported data, and when — one that cannot be quietly rewritten afterwards. Traceability is not only about production data; it also has to cover changes to the system itself, or the credibility of everything else is open to question. Also known as: Change log

Data residency

Keeping video and judgement data on the customer’s own servers or plant network so it never leaves the site. This is a hard requirement for many manufacturers bringing in an outside system, and it is a common reason edge deployment is chosen in the first place. Also known as: On-premise data

Video retention period

How long footage is kept before it is deleted automatically. Longer retention means better traceability and higher storage cost, so the period is normally set jointly from the customer’s traceability requirements and the rules of their industry rather than chosen by the supplier. Also known as: Retention period

Role-based access control

Granting view, configure and export permissions by job function. The credibility of quality data rests on a plain question: can anyone edit these records, or only people who are authorised to. Also known as: Access control

Term not listed?

Tell us the term you came across and we will add it to the glossary.

Contact Us

Follow a term further

• Choosing an industrial camera and lightingWhat shutter type, field of view, lighting and IP rating actually mean when specifying hardware.
• False alarm rates and alarm fatigueWhy recall and false alarm rate have to be read together, and how alarm fatigue gets a system switched off.
• Connecting Visual SOP to MESHow process validation results are written back to MES, and what each side has to prepare.
• FAQDeployment, accuracy and cost questions