Service · AI and ML engineering

Visual inspection with cameras

A camera and a model watching the line for defects, counts and safety, where a person cannot look all day.

Relevant to Manufacturing

The problem

A person inspecting a line all shift sees less by the sixth hour. A camera does not get tired, but out of the box it does not know your product.

Surface defects, missing components, label errors, fill levels, counts at a packing station, helmets in a yard: visual checks that are repetitive, continuous and recorded nowhere.

What it costs today

Typical: defects caught by the customer rather than the plant, a full-time inspector per line, and no record of what passed and what did not.

This is for you if
  • A check a person makes by looking, repeatedly, at the same kind of item or scene.
  • You can collect a few hundred examples of good and bad, or are willing to over a few weeks.
  • The line can take a camera, a light and a small compute box.

Probably not for you if

  • Defects that even your best inspector cannot see consistently. If people disagree on the label, a model will too, and we check that in the assessment.
What we build
  • A camera station and a model that flags defects, counts items or checks compliance, in real time, on the line.
  • A record of every item inspected, with its image, so a customer complaint can be traced.
  • Alerts to the operator and a shift report of what was flagged.
  • Everything running on a small device at the line, with no internet needed.
How we build itIn this order
  1. 01

    Fix the optics first

    Lighting, angle and camera choice decide most of the accuracy. We spend the first weeks here, before any model is trained.

  2. 02

    Collect and label with your inspectors

    Your inspectors label the first set. Where they disagree with each other, we find out why before training on it.

  3. 03

    Train, then test on a later batch

    Accuracy is measured on production from a later week, not the training week, and reported as misses and false alarms separately.

  4. 04

    Run in shadow, then act

    The model flags alongside the inspector for a few weeks. When the numbers hold, it takes the first pass and the inspector handles the flags.

  5. 05

    Keep it honest

    A sample is re-labelled every month. New products or lots are added to the set before the model sees them live.

Data you already haveThis is usually enough to start
  • Examples of good and defective items, or the ability to collect them
  • Your existing inspection records and rejection reasons
  • Product and lot schedules, to know what is on the line when
  • Photos from customer complaints
In practice

Illustrative example

Camera inspection that catches surface defects the end-of-line check missed

Inspectors check parts by eye at the end of the line, thousands an hour. Cracks and thread damage still reach the customer, and each complaint brings a sorting charge and a difficult call.

Read the case study
Typical engagement
Duration2-week assessment, then 8 to 12 week pilot on one station
TeamML engineer, hardware and integration engineer, your quality head
You receiveWorking inspection station, measured miss and false-alarm rates, image record, shift report
PriceFixed price, quoted after the readiness assessment; hardware extra
What you own at the end
  • Camera station, model and software, running at the line on your hardware.
  • The labelled image set and the accuracy report.
  • A written procedure for adding a new product or a new defect type.

Everything runs in your environment and is handed over with documentation. Nothing stops working if we part ways.

Questions we are usually asked

What does the hardware cost?

An industrial camera, lighting and a small compute box for one station is a modest fraction of the project. We specify it after the site visit and you buy it directly; we take no margin on hardware.

Will it work in our lighting?

We control the lighting at the station with a hood or a fixed lamp, because ambient light changes by hour and season. This is the most common reason vision projects fail, and we design for it first.

Can it stop the line?

It can signal the PLC to reject or stop, once the shadow phase has proved it. Most plants start with an alert and a reject bin.

How many defect images do you need?

Fewer than people expect for common defects, a few hundred. Rare defects are the hard part, and the assessment tells you whether you have enough.

Can the same cameras check safety compliance?

Helmets, zones and vehicle movement are the same kind of problem and often run on the same hardware. We scope them as a separate station so the quality work stays focused.

Often combined with

Not sure if your data is ready?

Take the free 20-minute readiness check. You get a short scorecard and a plain answer about what, if anything, is worth doing next.

Start the readiness checkTalk to us