Fraunhofer Institute of Optronics, System Technologies and Image Exploitation IOSB 
Circular flow of plastic packaging
As part of the BMBF funding measure "AI Application Hub Plastic Packaging - Sustainable Circular Economy through Artificial Intelligence", artificial intelligence (AI) is used to improve the sustainability of plastic packaging - along the entire value chain from packaging design to re-entry into the cycle. The K3I-Cycling project focuses specifically on increasing recycling with the help of AI so that more recyclates are reused as secondary raw materials.
Digital product passport for light packaging
K3I-Cycling aims for a quantitative and qualitative improvement of mechanical recycling of post-consumer plastic packaging waste. Therefore, K3I-Cycling is using an Artificial Neural Twin to develop a new, open and standardizable AI interface for the cross-sector collection of relevant information in the sense of a light packaging product passport. This will enable digital networking of all stakeholders along the value chain for the first time..
Our contribution
The Fraunhofer Institute of Optronics, System Technologies and Image Exploitation IOSB is researching hyperspectral imaging (HSI) methods for sensor-based sorting of plastic packaging. In addition to hyperspectral cameras in the short-wave infrared range already used for plastics sorting, sensors in the visible to near-infrared range and mid-infrared range are also being investigated for the various sorting tasks.
One focus is AI-supported, problem-specific wavelength selection. Efficient and economically attractive solutions are thus identified for specific sorting tasks (removal of impurities, sorting of plastic types, etc.). For the evaluation of the spectral data, the latest machine learning methods are used and optimized for the application. In particular, ensuring short inference times plays a key role here, as this is necessary to meet the real-time requirements of sensor-based sorting.
Another focus of the subproject is the investigation of approaches for model transfer. The goal is to develop methods that allow the use of previously learned classification and regression models even under changed boundary conditions (different camera type, change in illumination, etc.).
