www.simplify-data.de
Scratches on surfaces, missing components, misprinted labels: many production defects are visible to the naked eye, but only if someone looks very carefully for a very long time. Image recognition with artificial intelligence can support this visual inspection or partly automate it. This article explains how it works, what you need and where the limits are.
What computer vision does in quality control
Computer vision refers to methods that allow computers to analyse the content of images or video. In quality control, a camera captures the workpiece and software decides whether it passes or fails.
Typical tasks include:
- Defect detection: scratches, dents, cracks, discolouration or inclusions on surfaces
- Completeness checks: are all screws, connectors or labels present?
- Position and dimension checks: is a component in the right place, are distances and alignment correct?
- Reading codes and text: serial numbers, best-before dates, barcodes or QR codes
Rule-based machine vision or AI?
Industrial machine vision has been around for decades. Classic systems work with fixed rules: software measures edges, brightness values or distances, for example, and compares them against thresholds. This is very reliable when defects can be clearly defined and conditions stay constant.
AI-based methods, usually built on neural networks (deep learning), learn from example images instead. They are particularly useful when defects vary in appearance or are hard to express as rules, for example with natural materials such as wood or surfaces with changing textures.
| Criterion | Rule-based machine vision | AI-based image recognition |
|---|---|---|
| How it works | Fixed, programmed inspection rules | Learns from example images |
| Strengths | Precise measurement, easy to trace | Variable defect types, complex surfaces |
| Weaknesses | Inflexible when appearance varies | Needs many training images, decisions harder to explain |
| Adapting to new products | Program new rules | Collect new images and retrain |
In practice, both approaches are often combined: classic methods for measurement, AI for judging defects that are hard to describe.
What a project needs
Good images
The most important prerequisite is not the software but the image. Camera, lens and above all lighting determine whether a defect is visible at all. A scratch that stands out clearly under low-angle lighting may be invisible under frontal light. A well-designed inspection setup saves a lot of effort during training later.
Enough example data
An AI model learns from images of good and defective parts. Defect images are often the bottleneck: if a defect is rare, there are few examples. Some methods are trained only on images of defect-free parts and flag deviations from them (anomaly detection). They are a good option when defect examples are scarce.
A clear definition of defects
What counts as a defect, and what is still acceptable? If experienced inspectors cannot agree on a part, an AI cannot learn a clear decision either. A shared defect catalogue with limit samples is therefore a good starting point.
Integration into production
The inspection result has to trigger something: rejecting a part, raising an alert or being logged in a database. This requires interfaces to the machine control system or to existing IT systems. Cycle time matters as well: the analysis must be fast enough not to slow down the line.
Off-the-shelf system or custom solution?
Ready-made systems exist for many inspection tasks: smart cameras with built-in analysis software that can be configured through a user interface. They are quick to deploy and work well for standard tasks such as code reading or completeness checks.
A custom solution tends to pay off when the inspection task is unusual, when several cameras and data sources need to be combined, or when results must feed deeply into your own systems, such as an ERP system or Power BI reporting. Costs depend mainly on the inspection setup, hardware, data preparation, integration effort and ongoing operation. Flat prices are therefore hard to state reliably.
Limits of the technology
Computer vision is powerful, but it does not guarantee defect-free products. Keep these limits in mind:
- Error rates always exist: every system makes mistakes in two directions. It lets defective parts through (escapes) or rejects good parts (false rejects). Where to set the threshold is an economic trade-off.
- Visible defects only: a camera only sees what shows on the surface or in the image. Internal defects require other inspection methods.
- Changing conditions: new material batches, dirty lenses or altered lighting can degrade results. The system has to be monitored and retrained when necessary.
- New defect types: a model reliably recognises mainly what it has learned. Unknown defects may be missed.
- Traceability: for complaints or audits, you need to be able to explain why a part was released. Store inspection images and results.
Legal considerations
If employees also appear in the images, for example at a manual workstation, the GDPR applies. In Germany, the works council may also have co-determination rights when a technical system is capable of monitoring employee behaviour or performance (Section 87(1) no. 6 of the Works Constitution Act, BetrVG). Where possible, position the camera so that only the workpiece is in view.
Under the EU AI Act, pure quality inspection of components will in many cases not fall under the high-risk rules. This can be different if the AI system serves as a safety component of a regulated product, such as a machine. If in doubt, have such cases reviewed by a lawyer.
Checklist: how to start a computer vision project
- Define the inspection task: which defect on which product causes the most cost or complaints today?
- Record the current state: how is inspection done today, how much time does it take, which defects slip through?
- Create a defect catalogue: describe defects and set limit samples together with quality assurance.
- Collect images: capture them under real conditions with a good lighting setup.
- Test feasibility: before investing in equipment, check on a dataset whether the defects are detected reliably.
- Set target values: what escape rate and false reject rate are acceptable?
- Plan a parallel run: operate the system alongside the existing inspection at first and compare results.
- Secure operations: assign responsibility for monitoring, retraining and maintenance.
How Simplify Data can help
Simplify Data in Kiel develops AI solutions and software for companies in Schleswig-Holstein, Hamburg and across Germany. For computer vision projects, we support feasibility testing, software development and integration with existing systems and reporting. Learn more on our page about AI solutions for businesses.