HOLO-KI – Image processing with self-learning artificial intelligence


Short profile

Name: HoloKI

Involved: scitis.io, Werner Gießler GmbH, Fraunhofer IPM

Project duration/timeframe: January – December 2021

Brief description: With the help of self-learning AI-based image processing, holographically measured 3D data can be evaluated, thus reducing pseudo scrap in production.


Werner Gießler GmbH is a medium-sized company in the metalworking industry. The company produces high volumes of complex precision turned parts. Metalworking at high output rates requires individual inspection of each machined part, as defects as small as a few micrometers are often critical. Visual inspection using microscopes or magnifying glasses is still usually the solution today. However, this results in extremely high costs and there is potential for errors, so that high pseudo defect rates reduce productivity and increase CO2 emissions as well as resource consumption.

Since 2016, the Gießler company has been using a highly sensitive optical 3D inline measuring system – currently the fastest and most accurate in the world, developed at Fraunhofer IPM. This has already reduced the pseudo scrap from 30% to 10%. In order to further reduce the remaining 10% false reject rate, a time-consuming manual re-inspection of the supposedly poorly evaluated metal parts is necessary.


To solve this challenge, the existing holographic system was extended by an AI-based part classification. The goal was to detect the 10% pseudo scrap as such directly in the manufacturing process (inline) with the AI. The scitis.io algorithm examines the metadata from the holographic measurement for typical characteristics of a mismeasurement. The AI has to be trained permanently due to batch fluctuations and the possible changes of the pseudo error features. However, since this is not possible by means of manual labeling of image data, an additional strategy was developed that enables automated classification. By re-measuring borderline parts optically, the system can now permanently generate new training data of its own, which enables continuous improvement of the classification.

For this to work, the parts must be clearly assigned to the data from the initial measurement during a second measurement. This is done using a marker-free identification called Track & Trace Fingerprint, also a technology developed at the Fraunhofer Institute.


With the help of AI, around 98% of pseudo rejects can be directly identified as such. As a result, almost 1 million more components per year can be used that would otherwise be falsely declared as rejects. Rework is reduced by a factor of 100, which in turn increases productivity by 10%. In addition, power consumption and CO2 waste can be significantly reduced.

HoloKI - Holograph

Track & Trace Fingerprint

Tracing mass-produced components in the production history is important because even a small part, once installed, can reduce the functionality and longevity of the final product. But for small components, labels or other markings are too expensive or not even feasible. Therefore, there is the so-called Track & Trace process, which enables individual recognition. Many components have a microscopically individual surface structure or color texture. With the aid of an industrial camera, a defined area of the component can be photographed in high resolution. From this image, which depicts specific structures, a code number is calculated and assigned to an ID. This data is stored in a database with other product characteristics. For later identification, the process is repeated and the database is searched for the matching ID.

WordPress Cookie Notice by Real Cookie Banner