No Data – No Problem

Fraunhofer ITWM Generates Synthetic Data

Press Release / April 24, 2024

The »Virtual Inspection Planning« team at the Fraunhofer Institute for Industrial Mathematics ITWM and the BMW Group have successfully completed a joint research project investigating the use of synthetic image data, for development of automated surface inspection. This supports the inspection process using Artificial Intelligence and the standardized detection of defects. The project used synthetic image data to develop Machine Learning methods for inspecting highly specular metal surfaces. The experts, including representatives of the Fraunhofer ITWM and BMW Group from both Munich and Steyr InnoLab, presented their findings during a workshop at the BMW Group plant in Steyr, Austria.

Development of robust Machine Learning (ML) solutions requires a lot of annotated data, which is typically challenging to obtain. Synthetic data can remedy this shortcoming. In machine vision, this trend is particularly pronounced in scene-level applications (pose detection, object detection, motion detection, etc.). However, when it comes to detail-oriented applications such as visual surface inspection, it is much more difficult to generate synthetic data in a reliable and realistic way.

Synthetic Data in Visual Inspection

The aim of the project was to examine the usefulness of rule-based synthetic data generated by Fraunhofer ITWM in the development of automated surface inspection systems using Machine Learning. The rule-based approach to data synthesis is based exclusively on mathematical modeling and not on generative Artificial Intelligence and is integrated into the Fraunhofer ITWM's in-house software »V-POI«. This makes it possible to virtually simulate and expand inspection scenarios and to generate pixel-accurate and labeled data sets of any size.

Combining Joint Expertise

In the project, the BMW Group was responsible for developing the Machine Learning solution and carrying out a performance comparison between the real and synthetic data. The main task of the Fraunhofer ITWM researchers was to model surface textures and defects and to generate the photorealistic, synthetic image data set. The synthetic data generation parameters were adjusted to simulate the images obtained by the existing surface inspection system.

Pioneering Technologies in Surface Inspection

The study shows that the training results when using only synthetic images generated by the Fraunhofer ITWM synthesis pipeline are comparable to those when using only real images. The synthetic images contained acquisition scenarios that describe both the real setup and the extended edge cases. Furthermore, it became clear that training on the combination of real and synthetic data offers an unpredented advantage that even outperforms the results obtained when training only on real data.

© Fraunhofer ITWM
Comparison between a real image from the production line inspection system at the BMW Group (left) and a synthetically generated image generated using Fraunhofer ITWM technology (right).

Together, Fraunhofer ITWM and the BMW Group are advancing the field of machine vision for surface inspection by testing state-of-the-art synthetic data technology in a real industrial environment. Parametric control over data content, enabled by mathematical modeling and rule-based synthesis, is the future of AI-driven surface quality inspection. It offers possibility to make the inspection more robust, reliable and precise, even when availability of real data is limited.

Press Release

No Data – No Problem

Fraunhofer ITWM Generates Synthetic Data

Synthetic Data in Visual Inspection

More Information

Combining Joint Expertise

Pioneering Technologies in Surface Inspection

Contact Press / Media

Dr. Petra Gospodnetić

Contact Press / Media

M.A. Esther Packullat