Automated visual inspection engineered from the production constraint inwards: the time you have per part, the images you can actually get, and what happens to the parts it rejects.
Inspection projects are usually discussed as model problems. The decisions that determine whether the system works were made before any model existed: how many milliseconds there are per part, how many views that allows, whether the part presents consistently, and whether the lighting produces the same image every time.
Each of these is decided by the production environment rather than by the inspection team. Together they set the ceiling, and a project that starts with a model is solving the last problem first.
Two inspectors, two judgements
So the labels disagree before any model sees them.
Defects nobody has a photograph of
The rare ones are the expensive ones.
Scrap and rework costed in aggregate
So nobody knows what a false reject costs.
Inspection speed and coverage both limited
A sample at the pace a person can sustain.
Inspection at the end, not at the cause
Detecting late is better than not, and worse than early.
An override that operators have learned to use
Which is the failure mode nobody reports.
A pilot in a lab, on clean parts
Under lighting the line does not have.
The process changed and nobody told the model
New supplier, new lot, worn tooling.
Inspection results that are not traceable
So the record is a count rather than evidence.
Ask what a false reject costs you, as a number, for one product. Unit cost, the disposition of a rejected part, and the labour of reviewing it. Most plants cannot produce this figure, which means the detection threshold is being set by whoever configured the system rather than by the economics. It takes an afternoon with a cost accountant and it determines every tuning decision afterwards.
The sequence matters more here than on any other page in this section. Imaging comes before data, data comes before modelling, and the acceptance criteria come before all three, because they define what the labels mean.
We will also tell you when a project is not feasible at your line rate with your defect types, which is a more useful answer than a pilot that works on a bench.
Step 1
Assess the inspection process as it runs today
Step 2
Define the defect taxonomy with quality and production
Step 3
Establish acceptance criteria two inspectors apply the same way
Step 4
Assess the imaging: camera, lighting, presentation, vibration
Step 5
Establish the time budget from the line rate
Step 6
Cost a false reject and an escape, per product
Step 7
Collect and annotate images under production conditions
Step 8
Develop the model, with the imbalance handled honestly
Step 9
Integrate with the line: trigger, inference, result, actuation
Step 10
Pilot on the line, in parallel with existing inspection
Step 11
Validate against production conditions and agree thresholds
Step 12
Deploy with version control, monitoring and a retraining loop
Before committing to a vision project, run the inspector agreement check. Two inspectors, one hundred parts including the borderline ones, and count the matches. If they agree on ninety-five, the acceptance criteria are sound and the project has a specification. If they agree on seventy, the project is to write the criteria first. That is two days of work and it is the single best predictor of whether a vision programme will succeed.
Grouped by what is being inspected. Each is engineered from the line rate and the imaging constraints rather than from a model choice, and each is delivered with its false reject and escape performance stated.
Defect and surface inspection
Automated visual defect detection against a defined taxonomy with agreed acceptance criteria
Surface defect inspection: scratches, inclusions, contamination, finish variation, where lighting design does most of the work
Material and web defect detection on continuous processes, where the time budget is set by line speed
Defect classification as well as detection, where the type determines the disposition
Complex and composite defects, where the judgement depends on size, position and context together rather than on presence alone, which is where feasibility has to be established per defect type
Assembly and dimensional
Product assembly verification: presence, absence, position and orientation of components
Component presence and orientation checks, which are frequently the highest-value and lowest-difficulty cases
Dimensional and geometric inspection within the limits of what imaging can measure, stated honestly against your tolerances
Completeness and count verification, including kit and pack verification
Packaging, labelling and coding
Packaging and labelling inspection: correct label, correct position, legibility, seal integrity where it is visible
Optical character recognition and code verification: date codes, lot codes, barcodes and data matrix reading
Label content verification against the order or batch record, which is the check that creates the value rather than the reading itself
Where inspection results support product release, the validation strategy is determined by your quality and regulatory functions and we build and evidence to it
Platform and analytics
Edge inference deployment on hardware sized to the time budget
Line integration: trigger, result, reject actuation and interlocks, engineered with your controls engineers
Image storage, labelling and model lifecycle management, with version control at the line
Quality analytics on inspection output: defect trends by shift, material lot, tool and machine, which is frequently worth more than the inspection itself
Human review workflow for rejects, which supplies the labels for retraining
These are the common applications. The difficulty is rarely the model, and naming where it actually sits is more useful than describing the camera.
Automotive component defect detection
Surface and feature inspection on machined, cast or formed components at production rate.
Electronics assembly inspection
Component presence, orientation, polarity and solder condition verification.
Metal fabrication and surface inspection
Welds, finishes, coatings and surface condition on large or variable parts.
Pharmaceutical packaging inspection
Fill, seal, label and code verification at high speed.
Food and beverage packaging verification
Fill level, closure, label and date code verification, often at very high rates.
Textile and material defect detection
Continuous web inspection for flaws, contamination and pattern variation.
Counting and completeness verification
Confirming the right number and the right set of items in a pack or a kit.
Production line quality monitoring
Using inspection output as a process signal rather than only as a sorting mechanism.
Before the vision project, ask which defect type costs you the most per year. If the answer is not available by type, the plant is managing scrap in aggregate and cannot prioritise. That number also decides where to put the first camera, and it is frequently not at the end of the line but immediately after the operation that creates the problem.
We work within your existing line, controls and OT architecture. The architecture is determined by the time budget and the environment rather than chosen from a reference design.
Imaging
Cameras, optics and lighting selected for the defect and the surface, not for resolution alone.
Presentation
Fixturing and triggering, so the same view is captured every time.
Acquisition
The pipeline from trigger to image, deterministic within the time budget.
Inference
At the edge, on hardware sized to the budget, with no network dependency.
Result handling
Pass, fail, or uncertain, with uncertain routed to a person.
Line integration
PLC and SCADA signalling, reject actuation, interlocks, engineered with controls.
Records
Images, results and context stored to your traceability requirement.
Model lifecycle
Versioning, controlled deployment and a retraining loop.
Inference runs at the line. A production inspection decision does not depend on a network link or a remote service
Signalling to the control layer is engineered with your controls engineers, within your existing safety and interlock design, and reject actuation is a controls change rather than a software feature
Results flow to MES and your quality system so the inspection record is part of the batch or unit record rather than a separate database
Data flows outward from the line for analytics and retraining. Nothing inbound reaches the control layer
And the uncertain result is a first-class outcome. A system that must answer pass or fail on every part will be wrong more often than one allowed to ask
An inspection system makes a determination about product, which puts it inside your quality system rather than alongside it. The division of responsibility has to be explicit from the start.
Classification and standards
Defect taxonomy and acceptance criteria, owned by quality.
False reject and escape
Both measured, reported together, with the balance a business decision.
Validation
Against production conditions, to the strategy your quality function sets.
Human review and escalation
A defined route for an uncertain or disputed result, and an escalation path when the reviewer cannot resolve it.
Traceability
Result, image, context and model version in the production record.
Model control
Versioning and controlled deployment, per section 6.
Industrial security
Edge devices and network segmentation, nothing inbound to control.
Change control
Model, threshold and configuration changes through your process.
A vision model is trained on a production reality that then changes: a new material lot, a different supplier, worn tooling, a seasonal variation, a replaced lamp. The images shift and nobody tells the model.
Detection precision and recall by defect type, since aggregate performance conceals a type the model has stopped catching
False reject rate and escape rate, reported together and against the costed threshold, per section 7
Operator bypass rate, which is the trust measure and the earliest warning that the threshold is wrong
Inspection throughput and latency against the time budget, because a system that slows the line will be switched off whatever its accuracy
Image characteristics over time: brightness, contrast, focus and position, which detect a lighting or fixturing drift before the detection rate moves
Uncertain results as a proportion, which rises when the parts stop looking like the training data
Defect trends by machine, tool, shift, material lot and supplier, which is the analytics value from section 5
Scrap and rework against the pre-deployment baseline, which is the business case and needs the baseline to exist
Review outcomes on rejects, which supply the labels for retraining and should be captured as labels rather than as decisions
Model version in use per line, and the comparison record from each deployment
Time since last retraining, with the triggers rather than only a schedule
Monitor the images, not only the results. Brightness, contrast, focus and part position, trended. When a lamp ages or a fixture shifts, the images change before the detection rate does, which gives you days of warning instead of a quality incident. It is the cheapest monitoring on this page and almost nobody implements it.
Outcomes depend on your defect types, your line rates and the state of your acceptance criteria. These are the changes the work is aimed at.
The same standard is applied on every part, on every shift, which manual inspection cannot do
Defects are detected at the operation that creates them rather than after the value has been added
Scrap and rework fall against a baseline you measured before deployment
False reject rate is tuned against a costed number rather than against opinion
Operators stop needing the override, which is how you know the threshold is right
Every inspection result is traceable to an image, a context and a model version
The plant learns where its defects come from, by machine, tool, lot and supplier
The model improves from production data through a review loop that was designed in
And nobody can change what the line inspects without a change record
Bring us one inspection station and one defect type, and we will assess the imaging, establish the time budget, cost the error cases, and tell you honestly whether it is feasible at your line rate before anybody buys a camera.
CaliberFocus engineers visual inspection from the production constraint inwards: the time per part set by your line rate, imaging that is repeatable, acceptance criteria two inspectors apply the same way, and a costed false reject and escape. We deploy at the edge with version control at the line and a review loop that keeps the model improving.
- Imaging and acceptance criteria settled before any model is trained
- False reject rate and escape rate reported together
- Edge inference within the time budget, with no silent model updates
- Defect analytics by machine, tool, lot and supplier
Security & Compliance
