Visual defect detection and identification in continuous production processes minimal huma

Organisation/Company KU LEUVEN Research Field Computer science » Other Researcher Profile First Stage Researcher (R1) Final date to receive applications 5 Feb 2026 - 23:59 (UTC) Country Belgium Type of Contract Temporary Job Status Full-time Offer Starting Date 1 May 2026 Is the job funded through the EU Research Framework Programme? Not funded by a EU programme Reference Number BAP-2026-26 Is the Job related to staff position within a Research Infrastructure?

No

The project focusses on camera-based quality control, and more specifically defect detection and identification, in continuous production lines where goods such as polymer products, textiles, or food are transported on conveyor belts. The challenge in such processes lies in the nature of defects: they occur rarely, are often subtle, the underlying product may vary, and there is a wide diversity of possible defect types.

Developing such a defect detection system today requires lots of manual effort and custom product-dependent tuning that often starts with the collection of an offline dataset used to train models that are later put into production. These offline datasets often contain imbalanced data dominated by good samples, with many defect types underrepresented. Labeling such datasets is inefficient, as most labeled data will be defect‑free, and training models on these imbalanced datasets leads to poor performance.

Visual anomaly detection can offer a solution, but does not provide defect identification and is often hard to tune in practice. This project will focus on an autonomous inline system that starts without any prior knowledge and learns detecting and identifying defects on the fly. It will do so by analysing continuous incoming data, identifying and classifying deviations in the data while assessing the need for minimal human feedback to assist in identifying observed deviations.

The expected outcome will be validated both in the lab as well as in real-world industrial settings, hence requiring robustness and meeting all relevant industry standards.

The objective of this PhD is to contribute to the investigation, development, and valorisation of self‑learning camera‑based defect detection and identification systems for quality control in continuous production lines. This PhD position is part of the Flanders Make IRVA (Accelerator for Industrial Research and Valorisation) project RETINA, which intends to enable low‑effort visual defect detection and identification in continuous production processes.

It offers the opportunity to perform the research in close collaboration with leading industry partners.

• I have a Master's degree in Computer Science, Artificial Intelligence, Electrical Engineering, Mechanical Engineering or a related field and performed above average in comparison to my peers.
• I am proficient in written and spoken English.
• During my courses or prior professional activities, I have gathered experience with machine/deep learning, and can demonstrate a strong affinity with these fields. Prior experience with computer vision is a plus.
• I am proficient in Python and am familiar with data science and machine/deep learning toolkits. Experience with model deployment and the usage of MLOps tools (Dockerization, CI/CD pipelines, edge infrastructure, etc.) is a plus.
• As a PhD researcher at KU Leuven, I perform research in a structured and scientifically sound manner. I read technical papers, understand the nuances between different theories and implement and improve methodologies myself.
• Based on interactions and discussions with my supervisors and the colleagues in my team, I set up and update a plan of approach for the upcoming 1 to 3 months to work towards my research goals. I work with a sufficient degree of independence to follow my plan and achieve the goals. I indicate timely when deviations of the plan are required, if goals cannot be met or if I want to discuss intermediate results or issues.
• In frequent reporting, varying between weekly to monthly, I show the results that I have obtained and I give a well‑founded interpretation of those…