Computer vision and AI quality inspection
We build vision systems that inspect, count and monitor around the clock, using the cameras you already have wherever possible, and that flag problems before they reach your customer.
What computer vision does for a business
Computer vision is the branch of AI that lets software understand images and video. In practice that means a camera and a trained model doing a visual job a person does today: checking every part for scratches, counting cartons on a conveyor, reading a gauge, spotting a worker without a helmet, or measuring how long a queue is. A person can do this well for an hour. A model can do it consistently for every item, on every shift, and log every decision so you can see trends.
Most of our vision work falls into four groups:
- Visual quality inspection. Detecting surface defects (scratches, dents, cracks, discolouration, missing print), assembly errors (a missing screw, a wrong label, a reversed part) and dimensional problems on the line.
- Counting and tracking. Part counting, carton and bag counting, pallet tracking, vehicle counting at a gate, and stock levels on a shelf.
- Safety and compliance. PPE detection (helmets, vests, gloves), restricted-zone alerts, and fire or smoke detection from existing CCTV.
- Reading the physical world. Reading meters, gauges, number plates, batch codes and expiry dates, which overlaps with our document AI and OCR work.
Why manual inspection breaks down
Manual visual inspection is tiring, subjective and hard to audit. Two inspectors will grade the same borderline part differently, and the same inspector grades differently at the start and end of a shift. Sampling inspection catches only what is sampled. When a defect does escape, there is rarely an image to show what happened. A vision system addresses each of these: it applies the same standard every time, it can check 100% of output, and it stores an image and a decision for every item.
That does not mean replacing your quality team. The best results come when the model handles the repetitive first pass and routes uncertain cases to a person. Your inspectors then spend their time on judgement calls instead of staring at good parts. We explain the economics in detail in how computer vision cuts defects in manufacturing.
How we build a vision system
1. Define the defect, not the model
We start on your floor, not in a notebook. Together with your quality team we write down exactly what counts as a defect, with photographs of good parts, bad parts and the borderline cases that cause arguments. That defect catalogue becomes the specification the model is tested against.
2. Capture the right images
Image quality decides most of the outcome. We check lighting, camera position, resolution and line speed, and often add a simple enclosure or a diffused light to remove glare and shadow. Where your existing CCTV is good enough, we use it. Where it is not, we recommend industrial cameras that suit the budget.
3. Prototype on your data
We collect and label a few hundred to a few thousand images, train a first model and measure it against the defect catalogue. You see the detection rate, the false alarm rate and example images for each, before any production spend.
4. Deploy at the edge or in the cloud
Fast lines need decisions in milliseconds, so the model usually runs on an edge device next to the camera, which can trigger a reject gate, a light or an alarm. Results, images and dashboards sync to Google Cloud so managers can review trends from anywhere.
5. Keep it accurate
Products change, lighting drifts and new defect types appear. We monitor accuracy, collect the cases the model was unsure about, and retrain on a schedule so performance holds up after go-live.
What you get
| Deliverable | What it includes |
|---|---|
| Defect catalogue | Agreed definitions and reference images for every defect class |
| Trained model | Detection or classification model tuned for your parts, lighting and line speed |
| Edge or cloud runtime | Inference software, camera integration and reject-signal output |
| Review app | A simple screen where operators confirm or correct uncertain calls |
| Dashboard | Defect rates by shift, line, product and defect type, with image evidence |
| Support | Monitoring, retraining and an annual maintenance contract (AMC) |
Where it works well, and where it does not
Vision works best when the defect is visible in a photograph a person could judge, the part arrives in a fairly consistent position, and you can control the lighting. It is harder when defects are internal, when parts are highly reflective or transparent, or when there are only a handful of defect examples. We will tell you honestly in the first conversation whether your case is a good fit. Sometimes the answer is a simpler sensor, or better fixturing, rather than AI.
Industries we build vision for
Most of our vision work is in manufacturing (auto components, packaging, textiles, electronics, food processing), followed by logistics and warehousing (counting, damage detection, dock monitoring) and retail (shelf monitoring and footfall analytics). If you are new to the terms, our AI glossary explains object detection, classification and edge AI in plain language.
Frequently asked questions
How accurate is AI visual inspection?
It depends on the defect, the imaging setup and the amount of example data. Clear, visible defects under controlled lighting can usually be detected very reliably. We measure accuracy on your own parts during the prototype and share the detection and false-alarm rates before you commit to production.
Can you use our existing CCTV cameras?
Often, yes, for counting, safety and monitoring use cases. Fine surface inspection usually needs a dedicated camera and controlled lighting. We assess your current cameras during discovery and only recommend new hardware when it is needed.
How many defect images do you need to start?
A useful prototype can often start with a few hundred labelled images, including a few dozen examples of each important defect type. Where defects are rare, we use techniques such as anomaly detection that learn mainly from good parts.
Does the system need an internet connection?
No. Inspection models normally run on an edge device beside the line, so decisions happen locally in milliseconds. Internet is only used to sync results and dashboards to the cloud, and the system queues data if the connection drops.
How long does a computer vision project take?
A working prototype on your data typically takes a few weeks. Production deployment on one line follows once the accuracy targets are agreed. Timelines depend on image collection, hardware lead times and integration with your line.
What does AI quality inspection cost?
Cost depends on the number of cameras, line speed, defect complexity and integration work. We quote a fixed price for the prototype phase, then a build price and an annual maintenance contract. See our pricing page for how engagements are structured.
Have a process worth automating?
Tell us the one task eating your team's time, and we'll show you what AI does with it. We reply within one working day.