AI image analysis

AI image analysis uses machine learning models to analyse images automatically. In quality inspection, these systems can identify surface defects, deviations or anomalies without a person having to assess each image individually.

Unlike conventional, rule-based image processing, an AI model learns which patterns indicate a defect from a large number of example images. It can be adapted to new or changing defect patterns without having to reprogram every rule manually, although this may require further training with additional example images.

How does AI image analysis work in quality inspection?

The process generally follows three steps:

1. Image capture. Cameras capture images of the test object under uniform conditions. Consistent lighting and a fixed camera angle are crucial to keeping results comparable across different test runs. ‍
2. Image processing. Image contrast is enhanced and relevant areas are highlighted, often using thresholding methods or contour detection. This makes anomalies easier to identify.
3. ‍Classification. A trained model assigns the detected anomalies to a category and assesses the size, distribution or severity of the defect.

Practical example: detecting stone chips on paint surfaces

Stone chipping is damage to vehicle paintwork caused by the impact of small stones at high speed. Inspectors traditionally inspect the damage visually according to defined test standards. However, their experience and individual judgement influence the assessment. Different inspectors can therefore evaluate the same damage pattern differently.

To make the assessment more objective, a research team developed an image processing system that automatically analyses test panels following standardised stone chip tests. The system uses a sequence of contrast enhancement, thresholding methods and contour-based analysis to identify damaged areas based on their size and distribution.

The published results show that the automated analysis correlated strongly with the mean of the assessments made by experienced assessors. The individual test results, by contrast, differed significantly from one another. The automated inspection also took only around one-seventh as long as a manual analysis.

The example illustrates that AI image analysis is particularly well suited to identifying clearly defined, visually recognisable defect patterns that have so far required slow, subjective manual assessment.

What the method achieves

• Consistent assessment standards: A trained model applies the same criteria to every image. Declining concentration or fatigue can influence a manual assessment but does not affect automated analysis. ‍
• Speed: Large numbers of tests can be analysed in a fraction of the time needed for manual inspection, which is particularly relevant in series production.
• ‍Traceability: Each analysis can later be traced back to the underlying image.
• Detecting hard-to-see defects: A system trained for this purpose can also detect small or low-contrast defects that can easily be overlooked during visual inspection.

Limitations

AI image analysis assesses what is visible in the image. It does not explain why a defect arose. A specialist must still assess whether stone chip damage is due to insufficient coating thickness, an unsuitable formulation or a process fault.

For supervised AI models, the defect types the system should recognise must be defined in advance. The model also needs enough example images of each defect type. It may therefore overlook or misclassify rare or previously unknown defects.

The system requires careful calibration when the appearance of defects depends on lighting, the material surface or the camera angle. Without this calibration, the results are noticeably less reliable.

A single test result is an isolated data point. It only provides a robust basis for decisions when linked to experimental data, formulations and previous test results. For example, these combined data can determine whether a particular formulation is systematically more susceptible to stone chipping than another.

Frequently asked questions

Does AI image analysis completely replace visual inspection by specialists?

For clearly defined defect patterns, it can largely take over the initial assessment. Inspectors remain responsible for borderline cases and for assessing the cause of the defect.

Does every company need its own training data?

Usually, yes. Image quality, test conditions and the relevant defect patterns differ depending on the material and application.

How does AI image analysis relate to Material Intelligence?

Image analysis provides a single data point for a test object. That data point must be linked to additional material and experimental data to support well-founded decisions.