How computer vision cuts defects in manufacturing

The short answer

Computer vision reduces defects in two ways. First, it inspects every part instead of a sample, so fewer defective parts leave the plant. Second, it logs the defect type, time, machine and an image for each rejection, so your team can find and fix the cause. The first effect lowers customer returns. The second lowers scrap and rework over time. Both depend on good images and a clear definition of what counts as a defect.

If you want the service overview first, see our page on computer vision and AI quality inspection. This article explains how it works on a real line and how to judge whether it is worth doing.

Where vision inspection works well

Vision systems do best where the defect is visible, the part is presented consistently and the volume is high enough that manual checking is slow or patchy.

• Surface defects such as scratches, dents, burrs, paint runs and discolouration on castings, pressed parts, plastics and packaging.
• Assembly checks such as a missing fastener, a wrong label, a reversed gasket or a missing cap.
• Print and code checks such as batch codes, expiry dates and barcode quality.
• Counting and fill checks such as parts per tray or bottles per carton.

It does worse where the defect is internal (a weld void, a crack under a coating) and needs X-ray or ultrasound, or where the part is presented at random angles with no fixture. In those cases the answer is often to fix presentation first, then add a camera.

How the technology works, briefly

A camera captures an image each time a part passes a trigger. A trained model, usually an object detection or classification network, decides whether the image shows a defect and where. The decision can drive a reject gate, a light or an alarm, and the image and result go to a dashboard. For fast lines the model runs on an edge AI device beside the camera so the decision takes milliseconds and does not depend on your internet link.

The model learns from labelled examples of good and bad parts. That is why the first step is a defect catalogue agreed with your quality team, not a model choice. Our computer vision service starts there.

Manual inspection vs vision inspection

Factor	Manual inspection	AI vision inspection
Coverage	Often sampled	Every part, if the camera can see the defect
Consistency	Varies by person and shift	Same rule applied each time
Evidence	Usually none	Image and decision stored per part
Subtle or unusual defects	Strong with experience	Weak until examples exist
Cost profile	Ongoing labour	Upfront build, then lower running cost
Change handling	Retrain staff	Retrain model on new images

The best set-up is usually both. The model handles the repetitive first pass and sends uncertain cases to a person, which is a form of human-in-the-loop review. Your inspectors spend their time on judgement, not on staring at good parts.

What the line must provide

Most failed pilots fail on the physical set-up, not the AI. Check these before you spend money:

1. Stable lighting. Daylight changes and shiny surfaces cause false alarms. A shroud and a diffused light often matter more than the camera.
2. Consistent presentation. A fixture, guide rail or fixed conveyor position keeps the part where the camera expects it.
3. A trigger. A sensor or PLC signal tells the camera when to capture.
4. A place to act. A reject gate, a stop signal or at least a flashing light, so a detection leads to action.
5. Defect samples. Real bad parts, including the borderline ones. If defects are rare, plan to collect them over several weeks.

A worked example (hypothetical)

This is an illustration with made-up numbers, not a result from a client. Use your own figures.

Suppose a plant makes 50,000 parts a month. Say 1.5% reach the customer defective today, and each escaped defect costs Rs 800 in returns, replacement, freight and handling. Suppose the vision system, tested on your data, catches 70% of those escapes.

Formula: monthly saving = parts x escape rate x cost per escape x catch rate.

Input	Value
Parts per month	50,000
Escape rate	1.5% (750 parts)
Cost per escaped defect	Rs 800
Catch rate of the system	70%
Monthly saving	750 x Rs 800 x 0.70 = Rs 4,20,000

Now subtract the costs that come with it. Every false alarm sends a good part to rework or review, so if the system wrongly flags 1% of good parts (about 500) and each costs Rs 30 to recheck, that is Rs 15,000 a month. Add cloud and maintenance running costs. If the net monthly benefit were around Rs 3.9 lakh against a build cost of, say, Rs 12 lakh, payback would be about three months. If your escape rate is 0.2% instead, the saving falls to about Rs 56,000 and the case is much weaker. The point is to run this arithmetic before you build, not after.

Why a pilot beats a big rollout

Vision accuracy depends on your parts, your light and your defects, so a vendor demo on other people's images tells you little. A pilot on one station answers the real questions: what is the detection rate, what is the false alarm rate and how much effort does the review step add? It also shows whether your line needs lighting or fixture changes before you scale. See AI for manufacturing for other places a pilot can start, and the glossary entry on computer vision for terms used here.

What happens after go-live

A vision system is not a one-time install. Products change, a new supplier alters the surface finish, a bulb ages and the lighting drifts. Each of these can move accuracy. Good practice is to track the detection rate and the share of parts sent to review every week, to keep the images the model was unsure about, and to retrain on them on a schedule. We cover this under an annual maintenance contract (AMC), and it is the part buyers most often forget to budget for.

The other benefit appears after a few months. Because each rejection carries a timestamp, machine and defect type, patterns become visible: one die producing most burrs, one shift with more misprints, one supplier batch with a colour shift. Acting on those patterns is where scrap and rework come down. Without that follow-through, the camera only catches problems and does not prevent them.

How to get started

1. Pick one station with a clear, frequent, costly defect.
2. Photograph 50 to 100 good and bad parts under the lighting you have.
3. Write down the cost of one escaped defect and one false alarm.
4. Book a short call with us through the contact page and share the photos.

We then run a prototype on your images and report the detection and false alarm rates before you commit. Full details are on the computer vision service page. If you are still deciding whether AI fits at all, read build vs buy AI for SMEs or the AI readiness checklist.