Surface inspection with vicosys^® and AI

Reliable AI-powered fault detection from Vision & Control

for surfaces, containers and components throughout your entire production line.

The challenge in practice

Normal machining marks or genuine defects?

Not every irregularity on the surface is a defect. Tool marks, grinding marks, paint shadows or material texture are normal – traditional, rule-based systems often raise unnecessary alarms in such cases or overlook genuine defects amongst them.

A real-world example image: Are the fine lines and the material texture simply normal machining marks – or is there a fault hidden here?

The problem

Rigid thresholds and rules cannot distinguish between them: they treat every deviation from the ideal in the same way – regardless of whether it is inconsequential or actually a fault. The result is ‘pseudo-faults’ that cause good parts to be rejected, or missed defects that slip through.

The vicosys^® solution

The AI model learns from good images what the normal range of machining marks looks like and only triggers an alarm in the event of genuine deviations. This ensures that genuine defects are reliably distinguished from non-critical marks – without the need for manually maintained rule sets.

Benefits and Opportunities

High-resolution cameras

capture every detail of the surface in real time during the production cycle.

Telecentric lenses

provide distortion-free, true-to-scale images for reproducible test results.

Vision & Control System

synchronises the camera, lighting and signalling with the system’s inspection cycle.

vicosys^® Industrial PC

analyses images locally and in real time using AI algorithms.

Procedure

From the test pattern to the machine’s response

The camera captures an image – the test specimen is captured in time with the system’s cycle and sent to the evaluation computer.
vicosys^® evaluates the image – AI algorithms check the image for anomalies in real time.
Good part – cleared for further processing; the result is displayed on the monitor.
Defect – the signal lamp is activated, a stop signal is sent to the machine and a command is issued to eject the part.

AI tool – anomaly detection

Train the AI model, see the results straight away

The AI model is trained exclusively using images of healthy subjects. After training, the results are evaluated and the model is further refined. The sensitivity threshold can be adjusted using a slider. A heatmap visualises detected anomalies directly on the monitor.

AI tool – Segmentation

Recognising, distinguishing and counting objects

Automatic detection of different objects in the image – the AI segments and classifies each individual object.

Detection and counting of defect types – detected classes are reported individually, including area and number per type.

AI tool – Benefits

Easy retraining – new images can be taught at any time with no hassle.

Locally, without network access – the AI model runs offline, and sensitive image data remains with you.

Detects unknown types of faults – not all types of faults need to be known in advance; the model detects all types of anomalies.

Adjustable sensitivity – the model threshold can be easily adjusted using a slider.

Heatmap on the monitor – errors are displayed directly to the operator.

Flexible test object – suitable for surfaces, containers and components of a wide variety of shapes.

The Machine Vision Platform vicosys^®

ensures that your surface inspection is implemented as an industry-ready application.

High-performance algorithms, including AI, and integration of external libraries
Process-stable Linux operating system
Industrial communication (EtherCAT, PROFINET, I/O)
Can be combined with the entire Vision & Control standard portfolio of lenses and lighting systems

We would be happy to advise you on your bespoke inspection solution.

⇒ View the vicosys^® Machine Vision Platform product catalogue

Frequently Asked Questions

What distinguishes AI-based inspection from traditional, rule-based image processing?

Rule-based systems rely on pre-programmed thresholds and are therefore unable to reliably distinguish between normal machining marks and genuine defects. The vicosys^® AI model, on the other hand, learns from images of good parts what the normal range of variations looks like, and triggers an alert more selectively – only in the case of genuine defects.

How many high-quality images are needed to train the AI model?

The exact number depends on the complexity of the surface and the possible defect patterns. In practice, a manageable number of good images is usually sufficient to train a stable model – we will advise you on this on a case-by-case basis for your specific application.

Do I need to know all possible types of error in advance and tell the system about them?

No. As the model learns what constitutes a ‘good part’, it automatically recognises all kinds of deviations as anomalies – including defect patterns that were previously unknown or had not been specifically trained for.

How is the sensitivity of the error detection set?

The sensitivity threshold can be easily adjusted using a slider in the software, allowing you to optimise the balance between false positives and missed defects for your application.

Do our image files remain within the company, or are they transferred to the cloud?

The AI model runs locally on the vicosys^® industrial PC, without any need for network access. Your image data never leaves your company.

Can vicosys^® also distinguish between and count different objects or error types?

Yes. As well as simply detecting anomalies, the AI tool is also capable of segmentation: it automatically recognises, classifies and counts different objects or types of defects in the image.

What happens when an error is detected?

If an anomaly is detected, vicosys^® sends a signal to the PLC, which then controls the appropriate response of the machine.

Which camera and lens combinations are supported?

vicosys^® can be combined with the entire Vision & Control standard range of cameras, telecentric lenses and lighting systems, and can be configured to suit your test object.

Can the system be adapted retrospectively to new components or surfaces?

Yes, new images can be easily trained at any time, meaning that the model can be flexibly expanded even when test specimens or processes change.

For which sectors and test objects is this solution suitable?

The solution is suitable for surfaces, containers and components of a wide variety of geometries along the entire production line – from glass and metal surfaces right through to complex assemblies.