Skip to content
Neural Soft
Contact us Español (View this page in Spanish)

Use case

AI visual quality inspection

A camera and a computer vision model check every part, pack or label and set aside the defective ones. It inspects 100% of production and does not tire at the end of the shift.

  • Spots dents, cracks, stains, incomplete parts or misplaced labels.
  • Checks every unit, not just a sample.
  • Keeps an image of each defect so you can analyse why it happens.
Request an AI assessment See what it detects
A worker measuring a metal part by hand with a digital calliper.

Before and after

Manual inspection has its limits

Manual inspection

  • A sample is checked; the rest goes out unseen.
  • Attention drops as the hours and tiredness add up.
  • Two people do not always judge the same defect the same way.
  • There is little record to tell where faults come from.

With computer vision

  • Every unit is checked at line speed.
  • The criterion is the same at 7 in the morning and 10 at night.
  • Uncertain cases go to a person.
  • Each defect is logged with its image, time and batch.

Examples

What a camera with AI can detect

Machined parts

Image from camera 2

“Metal part on the conveyor, lit from above. There is a round mark of about 2 mm on the top face.”

Result

Status
Defect (confidence 0.91)
Type
Surface dent
Batch
L-2210

Part set aside for review

Bottles and caps

Image from the bottling line

“1-litre bottle with the cap slightly tilted and the seal not fully closed.”

Result

Status
Defect
Type
Badly closed cap
Action
Automatic rejection

Line supervisor alerted if it repeats

Labels

Image from the labelling machine

“Box with the label out of position and the batch code partly blurred.”

Result

Position
Off by 6 mm
Batch code
Unreadable
Best-before date
Correct

Box held back before palletising

How it works

From camera to decision, in a fraction of a second

  1. The camera takes the image

    With controlled lighting, at the right point on the line.

    1 image per part

  2. The model analyses it

    Trained on images of your good and defective parts.

    OK · defect · uncertain

  3. Action is taken

    A signal to divert the part, stop the line or raise an alert.

    Diverted to the review tray

  4. Human review

    Uncertain cases reviewed

    A person confirms the cases where the model is not sure.

    3 uncertain this shift

  5. Continuous improvement

    Confirmations are used to retrain the model and analyse root causes.

    Defect report per batch

Requirements

What it takes to work

Computer vision works very well in controlled environments. Before the project we check these three things.

  • Images of defects

    You need examples of good parts and of each type of defect. If defects are rare, they are collected over a few weeks or anomaly detection techniques are used.

  • Stable light and position

    A well-placed camera and constant lighting matter as much as the model.

  • A clear criterion

    What counts as a defect and what is acceptable must be defined by quality, with examples.

Automated manufacturing line with its control panel.

On the line

The camera fits into your line, without rebuilding it

The system is installed at a specific point on your current line, with its own lighting, and connected to your existing reject signal or a new one.

  • We choose the point on the line where the defect is easiest to see.
  • The trial runs in parallel, without touching production.
  • The shift's defects can be checked on a dashboard.

Method

We start with one defect and one point on the line

  1. Feasibility study

    We look at the line and the costliest defects, and take test images to check they can be detected.

    Feasibility report with real images

  2. Prototype

    We train a first model and run it alongside the current inspection, without touching production.

    Comparison with manual inspection

  3. Go live

    We integrate the camera and model into the line, with the reject signal and defect logging.

    Inline system and defect dashboard

Related service

  • Custom AI development

    We design and build artificial intelligence solutions tailored to your processes, data and tools when an off-the-shelf solution does not fit or falls short.

Related training

Which defects are slipping through today?

Tell us what you manufacture or pack, which defects are most common and how inspection works now. We will tell you whether computer vision can help.

Request an AI assessment