Quality Inspection System

A vision quality inspection system checks every unit coming off a line for defects, dimensions or missing components. Unlike a sampling inspector it never gets tired, never has a bad afternoon, and applies exactly the same standard to the first piece of the shift and the ten-thousandth.

Get a quote WhatsApp us

Who it's for

Manufacturing and packaging.

What changes

Defects caught at the station, not at the customer.

Starting at: ₹1,50,000
Timeline: 6–12 weeks
Category: AI Vision Systems
Built from: Vashi, Navi Mumbai

Key takeaways

Consistency, not raw accuracy, is the argument — human inspectors vary between themselves and across a shift.
Defect images are the project's critical input; 200–500 examples per defect type is the realistic starting point.
Lighting and fixturing typically cost more effort than the model does.
Set the system to over-flag rather than under-flag, then tune with production data.
Pilots run 6–10 weeks from ₹4,00,000; line deployment follows.

The honest case against human inspection

Human inspectors are good at spotting the unexpected and poor at repetition. Studies of visual inspection tasks consistently show accuracy dropping sharply after the first hour and varying widely between individuals looking at the same parts.

Sampling makes it worse. Checking one in fifty units means forty-nine defects can pass between samples, and the batch that reaches a customer is the one that was not sampled.

A camera checks every unit at the same standard. It will miss things a person would catch — anything genuinely novel — which is why the sensible design keeps a human on the line reviewing what the system flags rather than replacing inspection outright.

What can be inspected reliably

Not every quality check suits vision. This is roughly where the line falls in our experience across Indian manufacturing.

Inspection type	Feasibility	Typical accuracy
Presence or absence of a component	Straightforward	99%+
Surface scratches and dents	Good	92–98%
Print, label and barcode verification	Straightforward	99%+
Colour and shade consistency	Good with controlled light	95–99%
Dimensional measurement	Good with calibration	±0.1 mm typical
Weld and joint quality	Moderate	85–95%
Internal defects	Not viable optically	Needs X-ray or ultrasound
Texture and finish feel	Poor	Human judgement

Why defect images are the hard part

Every client has thousands of photographs of good parts and almost none of defective ones, because defective parts get scrapped rather than photographed.

The model needs examples of what it is looking for. Two hundred to five hundred images per defect type is a workable starting point, and collecting them is a project in itself — usually four to eight weeks of a supervisor photographing rejects as they occur, with defect type recorded.

We start this collection at the very first meeting, before any development, because it sits on the critical path and no amount of engineering compresses it. Where a defect is genuinely rare, we can sometimes augment with synthetic examples, but real images always outperform them.

Choosing the threshold, which is a business decision

Every inspection system trades false rejects against missed defects, and where you set that balance is not an engineering question.

If a missed defect reaches a customer and costs you a recall, set the system to over-flag and accept that a few good parts get pulled for human review. If your margin is thin and false rejects are expensive, tune the other way and accept a small miss rate.

We deliberately start deployments over-flagging. The first fortnight generates review work, but it produces exactly the data needed to tune correctly — and a system that starts by missing defects loses the quality team's confidence permanently.

Fitting into a running line

The system has to keep up with line speed and has to fail safely. If the inspection station goes down, the line should either continue with a flag or stop deliberately — never silently pass unchecked units.

Trigger. A sensor detects the part in position and fires the camera, rather than the camera guessing.
Decision within cycle time. Inference has to complete inside the time the part is at the station, which sets the hardware requirement.
Reject actuation. A pusher, diverter or simple light-and-buzzer, depending on line design and budget.
PLC integration. Where a PLC controls the line, the pass or fail signal goes through it rather than around it.
Traceability. Every image and decision stored against a batch, which is what makes a customer complaint answerable.

Running a pilot before committing

We do not quote a full line deployment from a specification. The pilot runs on one station, with one or two defect types, for six to ten weeks, and answers three questions: can the defect be seen reliably under achievable lighting, what is the real detection rate on production output, and what does the false reject rate cost.

From ₹4,00,000 for that pilot, including the camera, lighting, fixture and edge hardware for a single station.

If the answers are good, the line deployment is a known quantity. If they are not, you have spent a defined amount finding that out rather than committing to a system that was never going to work on that defect.

What you own afterwards

The inspection station hardware, the trained models, the image and decision database, PLC and line integration, a dashboard covering defect rates by type and shift, and full source.

The image database is the asset. Every inspected unit adds to it, and after a year it supports retraining that measurably outperforms the launch model. We build the retraining pipeline so your team can run it, because defects change as tooling wears and suppliers change.

FAQ

Quality Inspection System — your questions

How fast can it inspect?

Typically 5 to 30 parts per second depending on image resolution, the number of checks and the hardware. That is faster than most Indian assembly lines run, so line speed is rarely the constraint. Where it becomes one is high-speed packaging, where dedicated industrial vision hardware with hardware-triggered strobe lighting is the right approach rather than a general edge machine. We size against your actual cycle time at survey.

What if our defects are rare?

That is common and it makes the data collection harder rather than impossible. For defects occurring in under 1% of output, we sometimes invert the approach — train on what normal looks like and flag anything anomalous, rather than training on specific defect types. Anomaly detection is less precise but needs no defect images, and it works well as a first phase while real defect examples accumulate for a more targeted model later.

Can one system inspect multiple products?

Yes, with a changeover step. The system needs to know which product is running, either from the PLC, a barcode scan or an operator selection, and it loads the corresponding model and tolerance set. Where products differ substantially in size or shape the fixture may need changing too, which is a mechanical problem rather than a software one. Designing for changeover from the start is much cheaper than retrofitting it.

Do we need to connect it to the internet?

No. Everything runs on-premise on the edge machine, which is usually what factory IT prefers. An optional connection lets you view dashboards remotely and lets us support the system, but the inspection itself continues if the network is down. Production data staying inside the plant also removes a conversation about confidentiality that would otherwise involve your customers.

Who maintains it once you have handed over?

Your maintenance team handles the physical side — keeping the lens clean, checking the lighting has not degraded, verifying the fixture has not shifted. Those three things account for most performance drops we are called about. Model retraining is a periodic task your engineering team can run using the pipeline we deliver. We offer a support arrangement, but the system is deliberately built so you are not dependent on one.

Want a Quality Inspection System for your business?

Tell us what the process looks like today and we'll tell you what it would look like automated — and what it would cost.

Request a quote Talk to us