Skip to content

Quality control with computer vision

We build an image model that detects anomalies on the production line, working with your quality team. For Floating Farm, we installed the cameras ourselves.

For Floating Farm, we built an end-to-end solution, from installing the cameras to building the computer vision model

The model provides insight into the feeding process, helping Floating Farm reduce feed waste. It’s not a production line, but the chain is the same: an on-site camera, labelled images and a model that classifies in real time.

Read the case study

How many defects does our sample inspection actually miss?

A member of staff checks some of the products, but fatigue makes defects easier to miss. A camera checks every product in the same way and refers uncertain cases to someone with the right expertise.

Book a call

Every product checked, with uncertain cases referred to a person

Discuss your inspection process

A catalogue of defects

Together with the quality team, we define which defects matter: scratches, cracks, discolouration or a missing component.

Camera and lighting on the production line

The camera setup, lighting and timing of image capture matter more than the model. For Floating Farm, we installed the cameras ourselves.

A threshold that works both ways

Products that are clearly good pass through, those that are clearly defective are ejected, and the model refers everything in between to an employee.

Identifying defects in measurement data

For ProRail, we used machine learning to detect track components and wheel defects, making the railway safer.

We start with the defect catalogue; the camera comes next

  1. Defects and images

    Which defects matter, how often they occur and whether a camera can detect them on the line.

  2. Labelling and testing

    Your quality team reviews the images, and we test the model against a fixed set of known defects.

  3. Running alongside the line

    First, we observe without intervening. Then we let the model control rejections and feed the corrections back in.

Can a camera take over our quality control?

Explore it in a conversation

A good fit if

You visually inspect many similar products, defects are visible on the outside, and you want to check every product rather than a sample.

Not a good fit if

Defects are inside the product or only show up in a measurement, or they occur so rarely that there are no examples. In that case, a different sensor is a better fit.

What determines the scope

The number of types of defect, the speed of the production line, whether cameras are already installed, and whether products need to be rejected automatically or initially just flagged.

What if the model misses something?

We measure this in advance on a test set and keep track of it afterwards. A missed defect is fed back as an example so the model can learn from it.

Manual spot checks, fixed rules in a camera or a trained model?

Manual spot checks

Choose this if defects are rare and assessing them requires specialist knowledge. You accept that some products will not be inspected.

Fixed rules in a camera

Choose this if you check dimensions, presence or position under controlled lighting. When products or surfaces vary, fixed rules quickly reach their limits.

A trained image model

Choose this if defects vary in shape and conditions change. Our advice: start with one defect that currently takes a lot of time or leads to many rejections.

We’ve built image models that run on-site before

For Floating Farm and GVB, we built computer vision solutions that work with real camera footage. Data Science Lab has been operating since 2016, is a Microsoft Partner and Databricks Partner, and shares knowledge at Techdays.

Book an introductory meeting

We’ll start with the mistake that’s costing you the most right now

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your point of contact. In our first conversation, we’ll look at your production line, the errors and the images you already have.

Which anomaly would you like to detect automatically first?

Tell us what you check and how you do it now. We’ll reply by email or arrange an initial conversation.

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

A camera captures an image of each product, and a trained model checks it for defects you have defined in advance. The model gives an assessment and a confidence score. If it is uncertain, an employee checks the product.

Read about more uses of AI at AI applications.

That depends on how many types of defects there are and how different they look. Rare defects are the hardest because there are few examples of them.

That is why we start by collecting and labelling images, then count how many examples we have of each defect.

Yes. A model that learns what a good product looks like can flag anomalies without every defect being labelled. It can tell you that something is different, but not what it is.

We often combine this with classification for defects that occur more frequently.

Very important. No model can find a defect that is not visible in the image. We look at resolution, angle, lighting, reflections and the speed at which products pass by.

At Floating Farm, we installed the cameras ourselves before building the model.

If products need to be rejected immediately, the model often runs on a computer near the production line. If the aim is to report on results afterwards, it can run in the cloud.

Read about how we make that choice at developing computer vision solutions.

Using a fixed test set, we measure how many defects the model finds and how often it incorrectly rejects a product, broken down by defect type. Together with your quality team, we agree on an acceptable balance.

After launch, we continue to monitor this, including when products or lighting change.

The model may then perform less well. We monitor its performance and retrain it with new examples so it keeps up with production.

Read about how we manage models in production at MLOps.

For quality control, we point the camera at the product. If employees do appear in the images, we agree on what the model can see and what is stored.

For GVB, the model analysed camera footage without anyone viewing or sharing it.