Computer vision uses cameras and machine learning models so that a computer can "see" and decide: is this product defective, how many cartons are on the conveyor, is that worker wearing a helmet, what does that pressure gauge read. For small and mid-sized factories, the four uses below pay back within 6 to 12 months because they replace work that human eyes must do continuously and tire at. Success depends less on the model than on stable lighting, correctly labelled images, and someone operating the system after handover.
Why computer vision is now within reach of smaller factories
Three years ago a camera-based inspection system needed a dedicated engineering team and a budget in the billions of dong. Three things changed:
- Industrial and consumer cameras with sufficient resolution cost a few million dong.
- Pre-trained foundation models already recognise general objects; a factory only needs to teach them a few hundred to a few thousand images of its own products.
- Edge computers run models on the shop floor without sending images over the network, solving both speed and confidentiality.
The result: a single camera inspection point can now be deployed in a few months for a budget of a few hundred million dong, often less.
The 4 fast-payback uses
1. Visual defect inspection
Surface scratches, missing components, wrong colour, misaligned printing, open welds. This is the most common use because it directly replaces visual inspection, which is labour-intensive, inconsistent between day and night shifts, and prone to misses when people are tired.
Requirements: products pass a fixed position, lighting can be controlled, and there are at least a few hundred real defect images to learn from. Rare defects (a few cases a month) are much harder because there are too few examples; the alternative is to learn what a good product looks like and flag anything unusual, rather than learning each defect type.
Typical results: 60 to 90 percent fewer defects reaching customers, and 1 to 3 people per shift freed for other work.
2. Counting and traceability
Counting products on a conveyor, counting cartons onto a truck, reading codes on packages, matching them to production orders. Less glamorous, but a reliable payback because it replaces manual counts and end-of-shift reconciliation, a familiar source of inventory discrepancies.
The requirements are easier than defect inspection: object detection and motion tracking. Count data flows directly into the production or warehouse system; if the factory still runs on spreadsheets, see from spreadsheets to smart systems to prepare somewhere for the data to land.
3. Workplace safety monitoring
Detecting people without helmets or goggles, entering exclusion zones around running machinery, forklifts approaching pedestrians. On-the-spot alerts by light or siren, with events logged for review.
A point to consider: this is monitoring people. Be explicit with workers that the purpose is safety, do not store identity-recognising images when not needed, and set internal rules on who may review footage. Done properly, workers support it because it protects them; done covertly, it backfires.
4. Reading gauges, labels and screens on legacy equipment
Many older machines have mechanical dials or digital displays with no data port. A small camera reading the value every minute and writing it to the system is far cheaper than replacing equipment or fitting sensors. This is often the first step toward real operational data for predictive maintenance later.
What decides success is not the model
The shared experience of factory vision projects: the model is about 20 percent of the work. The rest:
- Lighting and camera position: 80 percent of accuracy problems live here. Light changes through the day, shadows from passing people, vibration from machines. A light-shielding enclosure costing a few million dong usually improves accuracy more than any model refinement.
- Correctly labelled images: the labellers must be people who know what a defect is by the factory's own standard. If two labellers disagree on the same image, the model learns that disagreement. This principle is the same as in any machine learning project; see machine learning fundamentals for business leaders.
- Thresholds and handling procedures: the model gives a defect probability; the factory decides the reject threshold and the send-for-recheck threshold. Where rejected products go, who rechecks them, how it is recorded, all must exist before the system is switched on.
- Operations after handover: products change design, suppliers change materials, cameras get knocked during cleaning. The model will drift. Someone must measure accuracy weekly and retrain when needed. This is the AI system management work that many projects skip before switching the system off six months later.
Real costs for one inspection point
Estimates for one camera position on one line, medium-sized products:
- Hardware: camera, lens, lighting, enclosure, edge computer: 30 to 120 million dong depending on line speed and required resolution.
- Build and training: image capture, labelling, training, alert and data integration: 100 to 300 million dong.
- Operations: 5 to 15 million dong a month covering accuracy monitoring, supplementary training, software maintenance.
Against the labour cost of three-shift inspection and the cost of defects reaching customers, most factories pay back within 6 to 12 months at the first inspection point; subsequent points are cheaper because infrastructure and procedures are reused.
An 8-week pilot plan
- Weeks 1 and 2: choose one inspection point, install camera and lighting, capture 2,000 to 5,000 images under real conditions across all three shifts.
- Weeks 3 and 4: label with the factory's best inspector, train the first model, measure on images not used for training.
- Weeks 5 and 6: run in parallel with human inspectors, record both results, no effect on the line. Compare, find the gaps, fix lighting or add images.
- Weeks 7 and 8: switch on real alerts with conservative thresholds, humans recheck cases the model is unsure about. Lock in the numbers: detection rate, false alarm rate, processing time per product.
By the end of week 8 there is enough data to decide whether to expand or stop, at a fraction of the cost of a plant-wide rollout.
5 reasons projects fail
- Starting with the rarest, hardest defect because it causes the most damage. Start with common defects to build data and trust, then move to rare ones.
- Capturing images in a lab, running on the shop floor. Different light, dust, vibration. Capture images at the exact position where the camera will live.
- No procedure for rejected products. The system flags a defect, nobody knows what to do next, workers silence the alarm.
- Evaluating by feel. Without detection and false alarm rates, every discussion about the system is opinion.
- Nobody accountable after handover. The system runs correctly for six months, the product changes design, accuracy declines gradually, nobody measures, the system is switched off.
Next step
Siri9 builds computer vision systems under AI integration, connects them to existing production software, and keeps accuracy stable under AI system management. If your factory is considering this, send us a description of the most labour-intensive inspection step and a few images of good and defective products; we assess feasibility within 3 working days.
Frequently asked questions
How many images are needed for training?
For common defects, a few hundred defect images and a few thousand good ones are enough for a first version. For rare defects, the learn-good-and-flag-unusual approach needs fewer defect images but has a higher false alarm rate at first.
Do images have to be sent over the internet?
No. The model runs on an edge computer in the plant; only aggregate figures and images of defect cases are stored centrally. This is both fast and keeps product images inside the factory.
Can it keep up with a fast line?
A mid-range edge computer processes 20 to 60 images per second with a typical defect model. Faster lines need high-speed cameras and specific optimisation; this is a point to measure concretely before committing.
How often does retraining happen?
When the product design changes, materials change, or weekly measurements show accuracy dropping. Factories with stable products typically top up training quarterly; factories that change designs often need a leaner retraining process from the start.