In-line visual inspection using unsupervised learning
| Reference number | |
| Coordinator | Linnéuniversitetet - Linnéuniversitetet Inst för datavetenskap och medieteknik |
| Funding from Vinnova | SEK 3 883 951 |
| Project duration | January 2023 - December 2025 |
| Status | Completed |
| Venture | Advanced digitalization - Enabling technologies |
| Call | Advanced and innovative digitalization 2022 |
Important results from the project
The project advanced knowledge on how AI can detect defects in industrial images, despite challenges with image quality. Normalizing Flows proved unsuitable in real factories, prompting a shift to other methods. The GLASS approach was further developed into a working prototype tested in industry and on edge hardware. The project also produced publications, open datasets, and collaborations that support future development in Swedish industry.
Expected long term effects
The project shows that academic AI methods often fail in industry due to data quality issues and limited validation on real datasets. Open data and new collaborations enable more robust methods. Long-term, the results can reduce manual work, improve quality, and offer new solutions to current limits in anomaly detection. Close cooperation between industry and academia is essential for future work in order merge fundamental research with practice.
Approach and implementation
Throughout the project, the consortium worked closely together, holding monthly meetings to share insights and address challenges. The researchers focused mainly on evaluating existing methods and developing new, effective solutions, while the industry partners collected higher‑quality data and tested and validated the methods in practice. Strong trust, high engagement, and the partners’ drive have been key success factors. Together, we make a difference!