AI image processing for highly reflective components in quality control

Background: Challenging surfaces, high requirements

At EVVA, metallic components undergo several process steps and must meet strict quality specifications. These include, among others, locking cylinders, cylinder cores, and other finely machined or galvanised precision parts from access technology. Mechanical defects such as scratches under 0.1 mm and galvanisation errors must be detected – deficiencies that are often hardly visible to the naked eye.

The central difficulty: Highly reflective surfaces reflect light and may conceal relevant features. For a reliable assessment, multiple viewpoints are often necessary. In practice, this complicates maintaining a consistently high quality of manual inspections. A solution was sought that reliably detects defects, supports employees, and integrates stably into the production process.

Solution: Closed test box, coordinated lighting, AI

To reliably solve these challenging inspection tasks, an AI-powered inspection system was developed in collaboration with Danube Dynamics. The system captures components arranged in a 400 × 300 mm Eurobox in one step, automatically assesses surface defects, and integrates quality control reproducibly into the production process. Up to 126 components are arranged in the box and are tested in one pass. To create constant conditions despite reflections, a closed test box with customised lighting is employed. Four individually controllable RGBW bars allow for different colours and lighting angles to make various defect patterns visible. The coordinated interplay of lighting, image quality, and AI evaluation was crucial.

The image capture is performed by a monochrome industrial camera of the type IDS uEye U3-36P0XCP-M-GL Rev. 1.2. It is based on the 19.80 MP sensor AR2020 from onsemi, delivers 5K UHD resolution with 5136 × 3856 pixels, and is suitable for inspecting the finest surface details. The 1/1.8-inch rolling shutter sensor in a 4:3 aspect ratio supports detailed imaging of even the smallest defects. The system is complemented by an IDS lens with a focal length of 8.5 mm (Type 20M11-C08528).

The camera captures the entire Eurobox in high resolution. The image data obtained in this manner enables the detection of the smallest defects in each individual component using AI subsequently. What is crucial is not only the pixel count but also the repeat accuracy under defined lighting conditions. The monochrome version additionally supports a high-contrast representation of relevant surface features.

“Especially with highly reflective surfaces, it is essential to depict fine differences stably and in detail,“ emphasises Jürgen Hejna, Product Owner 2D Cameras at IDS. “The camera's high sensor resolution and easy integration provide a solid foundation for industrial applications.“

Ease of integration was also a decisive factor for Danube Dynamics. The camera can be integrated into the overall system via a standard interface. This reduces the development effort and ensures stable testing results during ongoing operations.

Measurable benefit

The potential of the solution became evident during the first live deployment: The AI detected deviations that even experienced professionals hardly noticed. Above all, the significantly reduced testing time made the benefit immediately measurable.

The previously manual visual inspection of more than 30 seconds per Eurobox has been reduced to under five seconds. The basis for this consists of four defined lighting scenarios, from which four images per component are generated. The subsequent AI inference occurs with a runtime of approximately 10 milliseconds per cylinder.

At the same time, the testing process has been fully digitised. Testing results are now systematically documented and statistically evaluated. Scrap and deviations that were previously not recorded are now transparently traceable. Manual quality control is therefore not replaced but specifically supported. Employees are relieved, testing processes are reproducible, and quality is measurably integrated into the production process.

Another milestone is the ability to retrain the AI independently. EVVA can integrate new product variants and error patterns into the system by itself. This enhances the future viability of the solution and allows for adjustments to growing requirements without altering the fundamental system structure.