Computer vision · Defect detection systems
Defect detection systems that stop a flawed part from ever leaving the line
Banao builds defect detection systems that inspect every unit on your production line — surface cracks, chips, dimensional mismatches, contamination, and assembly failures — and signal your reject actuator before the part moves to the next station.
The system we deliver is the camera rig, the trained model, the edge hardware, the PLC integration, and the drift monitoring that keeps it honest long after go-live. It is tuned to your tolerance and your real cost of a miss, not a public benchmark.
RAK Ceramics— we delivered vision defect detection on a live ceramics line in the UAE.
The first call is free · 45 minutes · no obligation
What we build
What a defect detection system from Banao includes
A defect detector that holds up in production is more than a model. It is a trained system built on your own parts, with the image pipeline, the line integration, and the monitoring that keeps it from drifting when your line changes.
Surface crack and fracture detection
Hairline cracks, fractures, and surface breaks detected on every unit at line speed — the defects a fatigued inspector misses on the third shift, caught before the part moves on.
Chip, pit, and void detection
Material loss at edges, corners, or faces — chips, pits, and surface voids — detected in-line and logged with image evidence before they become a customer return.
Contamination and inclusion detection
Foreign matter on or embedded in the surface — oil, fibre, debris, inclusions — detected by texture and colour deviation from the clean-unit baseline the model was trained on.
Dimensional deviation flagging
Sub-pixel measurement of critical dimensions against your tolerance so an out-of-spec part is rejected on the line, not at a customer's incoming inspection.
Colour and gloss deviation
Finish defects — uneven colour, gloss banding, staining, oxidation — detected against a calibrated reference so a visibly off unit does not reach a box.
Texture and pattern anomaly detection
For lines where defect classes cannot be fully enumerated, models trained on good units alone flag any unit whose texture or pattern deviates from what correct looks like.
Dataset build and active-learning loop
We capture, label, and curate images from your own line and build the active-learning loop that keeps feeding the model as new defect variants appear — the model improves over time instead of freezing at go-live.
Threshold tuning to pass/reject economics
We set the detection threshold against your actual cost of a miss versus a false reject, not a generic accuracy score, so the system operators see is one they can trust.
Why defect detection pilots stall before they reach the line
The pilot that impressed your quality team ran on curated images, good light, and a defect rate high enough to show results clearly. Your production line gives you vibration, shift changes, a new material batch that changes how a surface reflects, and real defect rates in the parts-per-thousand. Moving from the first condition to the second is where most detection projects stop.
We develop for the production condition from day one. Controlled presentation — consistent lighting, correct optics, stable part orientation — comes before model training. A labeled dataset from your own line comes before acceptance testing. A threshold set against your pass/reject economics comes before deployment. The model trains in a week; the rest is the project.
Presentation before model
We fix the image before writing training code. Consistent lighting and optics make a small model reliable; an inconsistent image defeats the best model available.
Threshold tuned to your economics
The correct threshold is the one that minimises your real cost — customer returns, rework, and line stoppages — not the one that maximises a benchmark score. We set it against that.
Validated on your held-out parts
Before go-live we hold the system to a written target on detection rate and false-reject rate, scored on a held-out set of your own parts — not a public dataset whose defects are not yours.
False-reject rate tracked from day one
A system that catches every defect but also rejects good units gets switched off by operators within days. We track false-reject rate as a primary output, not an afterthought.
What keeps a defect detection system accurate at month six
A model frozen at go-live is a model in slow decline. A new material supplier changes how a surface reflects. A tool wears and the part geometry shifts. A lighting tube ages and the image dims. Six months later the detection rate has dropped and nobody noticed, because the monitoring was not there.
We build drift monitoring and a retraining path in from go-live, not as a later-phase addition. The system that passed acceptance and the system that works after a season change are the same system only if drift is caught and corrected — not accepted as normal decline.
Detection rate and false-reject rate in production
Dashboards on both numbers. A rising false-reject rate is the leading indicator that something changed; we alert on it before operators start overriding the system.
Image-quality monitoring
Blur, underexposure, and contrast loss in the input image degrade a model before the defect statistics tell you. We monitor image quality as a leading signal.
Active-learning retraining loop
When a new defect class appears or the material changes, the loop captures and labels the new examples and retrains the model — designed in from the start, not an emergency response.
Receipts
Defect detection already running on real lines
Metrics shown dotted (··) are being verified in our case-study metrics pack. The deployments are live.
Surface defect detection on a live UAE ceramics line
We deployed vision defect detection on a ceramics production line in the UAE, classifying tile surface and finish defects at line speed and driving the reject gate before packing. A named client running in production in the region we have a direct presence in.
Crack and surface defect detection on a machined-parts line
Vision detection of cracks, chips, and surface damage on machined components, running in-line with the verdict fed to the existing PLC reject gate. Every detection logged with the image for traceability and warranty investigation.
Dogfooding
We measure our own AI the same way we measure yours
Banao runs InterviewGod and Vikaas — AI systems we built — on our own company, in production, every working day. They have to be right on real inputs. We monitor them for drift, retune them when they slip, and would switch them off if they stopped earning their place. That is not a separate discipline from how we build a defect detection line; it is the same standard applied to a camera and a model instead of a language system.
A detection system that passes acceptance and then quietly degrades is not a working system. The monitoring, the retraining path, and the human gate on uncertain verdicts that we build into your line are the things we hold our own AI to.
AI we run on our own hiring — held to a measured accuracy bar, every week.
AI we run on our own demand generation, monitored in production daily.
The honest version
When a defect detection system is not the right answer
Vision detection is not the right tool for every quality problem. We would rather tell you on the first call than discover it partway through a build.
- The defect is not visible on the surface: internal voids, sub-surface cracks, or material composition failures need X-ray, ultrasound, or eddy-current. A camera cannot see them, and we will say so.
- Your defect rate is too low to validate: if real defects are so rare you cannot hold out enough examples to score accuracy, neither you nor we can trust a detection rate claim.
- Natural variation overlaps the defect signature: some materials vary legitimately in ways that resemble the defect. If the two cannot be reliably separated, detection will reject good units and get switched off.
- A simpler sensor already solves it: a weight check, a laser gauge, or a contact sensor can outperform vision on reliability and cost for the right defect. We will point you to it.
- Presentation cannot be controlled: if parts arrive in random orientation under uncontrolled light, and that will not change, the line needs fixing before a model is trained.
How we start
How we start — test the hardest defect before building the system
We prove feasibility on your hardest defect class before asking you to commit a build budget. If your last pilot did not make it to the line, this is why.
- 01
AI Discovery Sprint
2 weeks · fixed price
Bring samples or images of your hardest defect class. We test whether a model can detect it under your line conditions, then hand back a feasibility verdict, a rig and integration plan, and ROI maths — yours to keep either way. If you proceed, the Sprint cost is credited against the build.
- 02
Build and integrate
We design the camera and lighting rig, capture and label the dataset from your line, train and validate to your acceptance criteria, and wire the detection verdict into your PLC, reject gate, and quality records.
- 03
Production and drift monitoring
We deploy at line speed with monitoring on detection rate, false-reject rate, and image quality. The active-learning loop captures new examples as the line changes and keeps the model on target.
FAQ
Frequently asked questions
What is a defect detection system?
A defect detection system uses cameras and trained vision models to inspect every unit on a production line for specific defects — surface cracks, chips, contamination, dimensional errors, colour deviations — and to send a pass or reject verdict to the line automatically, at production speed.
How does AI defect detection differ from rule-based machine vision?
Rule-based machine vision uses fixed thresholds: brightness above a value, an edge in the expected place. AI defect detection learns from labeled examples of good and defective parts, so it handles the natural variation in real materials and catches the irregular, multi-cause defects that rules miss. The tradeoff is that you need labeled examples and a validation set — which is why the data pipeline matters as much as the model.
How many defect images do you need to train the system?
Fewer than most projects assume. For common, visible defects a few hundred labeled examples can be enough to start. For rare defects we use anomaly detection trained mostly on good units, plus augmentation. Either way, we build an active-learning loop that keeps adding real examples from your line over time.
Can the system distinguish a real defect from natural material variation?
This is the threshold-tuning question, and it is the core of the engineering work. We set the detection threshold against your actual pass/reject economics — cost of a miss versus a false reject — not a generic accuracy score. If natural variation and the defect signature overlap too much to separate reliably, we will tell you that before you commit a build budget.
What happens when the model is uncertain about a unit?
Units below a set confidence threshold are routed to a human inspector rather than silently passed or rejected. The confidence gate is agreed with your quality team, and every low-confidence decision is logged with the image so the pattern can be investigated and the model improved.
How do you stop accuracy from declining after go-live?
We monitor detection rate, false-reject rate, and input image quality in production and alert when any of them move. When a new material batch or a worn tool shifts the distribution, the active-learning loop captures new examples and retraining brings the model back on target. Drift monitoring is built in from launch, not added in a later phase.
What types of defects can the system detect — and which can it not?
Strong on anything visible at the surface: cracks, chips, scratches, pits, contamination, gloss and colour deviations, dimensional errors, missing or misplaced components, and wrong or unreadable labels. Not strong on internal defects: sub-surface cracks, voids, or material composition require X-ray, ultrasound, or other sensors, and we will tell you when that applies.
How long from first scoping call to a live detection system?
A common path: a 2-week Discovery Sprint to validate feasibility and size the build, then roughly 8–12 weeks to build the rig, label the dataset, train and validate, integrate with the PLC, and complete acceptance testing. Banao's engineering bench means work starts in weeks.
Get started
Bring the defect that keeps reaching your customers
Show us the check that costs you the most in returns, rework, or a shift of inspectors. In 45 minutes we will tell you whether a detection system can catch it — and what putting one on your line would take.
Book a Discovery Sprint