A vision system can inspect multiple product variants within the same inspection area, provided stable optical conditions are maintained. Adaptive lighting scenarios, software-side recipes per variant, and deep-learning models for irregular textures are decisive. The THOPinspect platform resolves variant changes through automated recipe matching directly via the machine control.
Why variants make inline inspection challenging
A scratch or surface defect on a high-gloss polished aluminium part behaves optically very differently to one on matte black injection-moulded plastic. While glossy surfaces reflect light directionally (specular reflection), matte materials scatter light diffusely in all directions. This significantly affects contrast transitions, thresholds and grayscale distributions on the camera sensor. Colour changes also distort brightness values in classic monochrome camera systems. A robust design must therefore never be based on an ideal “golden part”, but must include representative good and defective parts across all product variants.
Optical stability as the primary lever
Before image-processing algorithms or AI models come into play, the mechanical lighting concept must optically isolate defects. Typical industrial approaches include:
- Diffuse dome or coaxial lighting: Eliminates hard direct reflections and is excellent for strongly reflective surfaces or curved geometries.
- Dark-field ring lighting: Light strikes the object at a shallow angle. Defects such as scratches, chips or edges break the light into the camera and appear bright, while the defect-free surface remains dark.
- Structured LED lighting: Projects line or diamond patterns to geometrically capture three-dimensional shape deviations (e.g. dents or sink marks).
Software strategy: separate recipes versus hybrid models
In THOP AG's industrial practice, two approaches have become established depending on the degree of variance – more on this in our article on rule-based image processing vs. AI:
- Classic rule-based multi-recipe system: If the variants differ clearly in geometry or colour, the THOPinspect software stores a dedicated “recipe” per part. Thresholds, measurement tools and tolerance zones are statically predefined. When switching part types, the system loads the parameters via the PLC flawlessly within milliseconds.
- Learning AI models (anomaly detection): Where good parts show natural, process-related variation (e.g. differing fibre orientation in composite materials or varying casting structures), rigid thresholds reach their limits. Here, THOP AG trains a neural network exclusively on the good state, in order to detect unexpected deviations across variants.
If an existing system is being retrofitted rather than newly built, additional framework conditions apply – see our detailed article on retrofitting existing vision systems.
Use automated type verification for maximum process stability. THOPinspect systems validate at recipe change via a pre-check image whether the physically loaded part matches the recipe reported by the PLC, ruling out costly tooling damage or false inspections.