Visual Quality Inspection and Defect Detection
Automated inspection engineered from the lighting and optics upward, tuned to the two numbers your quality team actually argues about: how many defects escape and how much good product gets rejected.
Most failed inspection projects failed at the imaging stage and blamed the model. If the defect is not clearly visible in the image, no amount of training will find it, and the fix is a light, a lens or an angle rather than an architecture.
Visual quality inspection is the automated detection of defects, deviations and non-conformities in manufactured or processed goods using cameras and machine learning, integrated with line control so that a detection results in a rejection, a diversion or an alert within the cycle time available.
Imaging before modelling
The first fortnight is spent on the physics, because it determines everything after it:
- Lighting. Diffuse, directional, backlit, coaxial or structured, chosen for the defect type. A scratch invisible under flat light is obvious under grazing illumination.
- Optics and resolution. Enough pixels across the smallest defect you must catch, with the depth of field your part variation requires.
- Presentation. How consistently the part appears: position, orientation, vibration, and whether every surface that matters is actually seen.
- Timing. Trigger, exposure and cycle time, so images are captured sharply at line speed rather than blurred at it.
We prove the defect is visible in an image before any model is trained, with a sample of known good and known defective parts. Where it is not, we say so, and the recommendation is an imaging change rather than more data.
How we build inspection
Agree the defect taxonomy with QA
What counts as a defect, what counts as cosmetic, and what the tolerance is for each, written down with the people who currently make that judgement. Inspectors disagree with each other more than anyone expects, and that disagreement becomes the model's confusion if it is not resolved first.
Use anomaly detection where defects are rare
Well-run lines produce very few defects, which is a problem for supervised learning. Training on normal parts and flagging deviation works with the data you have, and it catches defect types nobody has seen yet. It also produces more false alarms, which is a trade we make explicitly.
Tune to escape rate and false reject cost
The threshold sits where the cost of a defect reaching a customer balances against the cost of scrapping good product. Both numbers come from your quality and finance teams, and the operating point is a business decision we document rather than a modelling choice we make.
Run at line speed, on the line
Inference within cycle time, on hardware that survives the environment, with a defined behaviour when the system is unavailable so the line does not stop. This is edge deployment with industrial constraints, covered further under optical inspection for edge devices.
Show operators why
A highlighted region and a defect class, not a score. Operators need to verify the call quickly, and a system they cannot check gets bypassed within a fortnight, usually by a supervisor under production pressure.
Capture every decision for traceability
Image, decision, confidence, operator override and outcome, retained to your policy. This is the record your quality system, your auditors and your customers will ask for, and it is also the training data for the next version.
Run in parallel before you let it reject anything
The system inspects alongside your current process for several weeks, and its calls are compared against the inspectors'. Disagreements are reviewed by QA and are usually the most informative part of the project, because a meaningful share of them turn out to be cases where the machine was right.
What it changes
| Measure | Manual inspection | After automated inspection |
|---|---|---|
| Coverage | Sampled, or 100% with fatigue effects | Every part, consistently, at line speed |
| Consistency | Varies by inspector, shift and hour | Identical criteria applied every time |
| Escape rate | Known approximately, from returns | Measured directly against parallel review |
| Traceability | Paper or spreadsheet records | Image and decision retained per part |
| Root cause | Anecdotal, after a batch fails | Defect type trends by line, shift and material |
The last row is frequently worth more than the inspection itself: defect trending by shift, batch and machine turns quality from a gate into a feedback loop.
How the engagement runs
Imaging is proven first, then the system runs in parallel before it is allowed to reject anything.
Imaging trial
Lighting, optics and presentation tested on known good and defective samples until the defect is clearly visible in the image.
Defect taxonomy and dataset
Classes and tolerances agreed with QA, inspector agreement measured, images collected across shifts and materials.
Model development
Supervised and anomaly approaches trained, escape rate and false reject measured against the agreed operating point.
Line integration
Inference at cycle time on line hardware, control integration, operator interface and fallback behaviour.
Parallel running and handover
Inspection alongside the current process with disagreements reviewed by QA, then graduated go-live and handover.
What you receive
An inspection system your quality team trusts, with the traceability your customers ask for.
Imaging specification
Lighting, optics, mounting and trigger design proven on real parts.
Defect taxonomy
Classes and tolerances agreed with QA, with inspector agreement measured.
Inspection model
Supervised or anomaly based, deployed on line hardware within cycle time.
Operating point analysis
Escape rate against false reject rate, with the cost basis for the chosen threshold.
Operator interface
Highlighted defect regions, class and override capture.
Traceability record
Image, decision, confidence and outcome per part, retained to policy and queryable.
Is this the right engagement?
Worth being direct. Visual Quality Inspection and Defect Detection is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- Inspection is manual, sampled, or inconsistent between shifts.
- The defect is visible to a person in a photograph, or imaging can be changed so it is.
- Line access is available for imaging trials and parallel running.
- QA can agree defect classes and tolerances.
- Escape and false reject costs can be estimated.
Choose something else if
- The defect is internal or not visually detectable, which needs a different sensing modality.
- The line cannot be accessed for trials or camera mounting.
- Cycle time leaves no window for image capture at usable quality.
- Nobody will act on detections, or the line has no reject mechanism.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
How many defect images do we need?
Fewer than expected if anomaly detection suits the case, because it trains mainly on normal parts. Where supervised classification is needed, a few hundred examples per defect class is a reasonable starting point, and rare defects are usually best handled by the anomaly path rather than by waiting to accumulate examples.
What if our defects are extremely rare?
That is the normal situation on a well-run line and it is why anomaly detection exists. The model learns what normal looks like and flags deviation, which also catches defect types nobody has seen before. The trade is more false alarms, which we tune against your reject cost.
Can it run at our line speed?
That is determined in the imaging trial, not assumed. Capture must be sharp at line speed and inference must complete within cycle time on hardware that survives the environment. Where it cannot, we say so early, and the answer is usually optics and lighting rather than a faster model.
Will it replace our inspectors?
In the engagements we run it changes what they do: the system inspects every part and inspectors handle flagged cases, adjudicate disagreements and work on root cause. The defect trending the system produces is frequently worth more than the inspection itself.
How do we know it is better than what we do now?
By running in parallel for several weeks and comparing calls against your inspectors, with QA adjudicating the disagreements. That comparison is the most informative part of the project, and it regularly shows the machine was right on cases everyone assumed were errors.
Often paired with this
Most clients combine two or three engagements from the Computer Vision pillar. These are the ones that most often run immediately before or after.
Optical Inspection for Edge Devices
Vision models running on device within real thermal, power and latency budgets, with fleet updates handled.
Read more →Object Detection and Image Classification
Detection and classification models trained on your own images, with honest labelling and cost-weighted thresholds.
Read more →Pose Estimation and Activity Recognition
Movement and activity understood over time, using skeletal data that needs no identifiable footage.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.