Skip to content

Case Study: AI-based quality control of decorative elements

Car keys undergo an extensive inspection for possible decorative defects, including, for example, scratches on the surface, air pockets, faulty castings or imprints (Photo: HELLA Aglaia)

At the end of a production process, end-of-line inspections are essential to ensure compliance with a constant and reliable product quality. Particularly in the case of an optical quality control, this often requires a large number of trained personnel. With the help of smart image processing based on artificial intelligence (AI), HELLA Aglaia enables reliable 100-percent inspections even in difficult cases, thus helping to increase automation in quality control.

 

Area
  • Production
  • Cross-Industry
Task Automation of a quality inspection
Challenge
  • Variable shapes and textures
  • Wide range of test conditions
  • Variable defect size, shape and orientation
  • Hardly visible errors
Solution AI-based detection of surface defects based on grayscale images with multiple directions of exposure
Result
  • Automatic end-of-line inspection for decorative defects
  • Error classification and localization
  • Simplified downstream root cause analysis for process optimization

Challenge

When a new vehicle is delivered, every detail must be right – even the vehicle keys must be free of any defects before they are handed over to the new vehicle owner. The keys therefore undergo an extensive inspection for possible decorative defects, including, for example, scratches on the surface, air pockets, faulty castings or imprints, or even missing individual parts. Detecting these defects, no matter how small, is costly and, above all, time-consuming. This is largely due to the detailed design of the keys, which includes different materials and variants.

Solution

A grayscale camera is used to visualize the defects. Several LEDs from different directions provide sufficient light and enable the lighting conditions necessary for visualization from several directions. In this way, different lighting scenarios can be generated and the various surface defects can be mapped reliably. All vehicle keys go through this camera-light setup at the end of production. In each case, several images are taken with different exposure scenarios and then processed with an optimized object detection algorithm – all within the line cycle time. The special feature of this image evaluation: No objects are detected, the algorithm only evaluates the detected defects as such. This skill was previously trained with a large number of annotated sample images from the same shooting setup. The algorithm is thus able to reliably detect the defect types represented in the training data on the different surface and material conditions – practically as if they were separate objects. Accordingly, the keys are classified as In order or Not in order and defective copies are either sorted out or assigned to rework, depending on the type of defect.

Results, Outlook

The implementation of this surface inspection demonstrated that an object detection approach can be applied to reliably detect, locate, and classify defects on decorative elements. Combining the localization and classification of defects also opens up options for root cause analysis: By assigning each defect, including the position and type, to the serial number of a key, the most common defects can be identified in downstream analyses and the causes eliminated. Automatic decorative end-of-line inspection is thus an integral part of process optimization for the entire production chain.

Further information