Description
Visual anomaly detection in an industrial setting requires reliably identifying deviations (defects) from the distribution of nominal samples (i.e., defectless samples) using optical inspection systems. This project focuses on the optical quality control of industrial steel bands in steel rolling mills. The assessment is crucial for evaluating the overall product quality and identifying emerging production issues. In particular, high reliability and robustness, e.g., against new lighting conditions or altered camera positions are essential. In addition, quick adaption to new products with only a small number of nominal samples is desirable.
|
The project aims to explore the adaptability of a recently proposed method, AnomalyDINO [1], to the domain of steel band surfaces. The method itself is based on a nearest-neighbor approach in the feature space of DINOv2, a pretrained vision transformer. The initial phase focuses on a Proof-of-Concept, assessing whether the method can be effectively adapted to this new domain. This involves identifying necessary adjustments to the model and the data pipeline, such as preprocessing, fine-tuning or substituting the backbone, and the task-specific scoring of anomalies. A refined (e.g., multi-stage) approach to object and anomaly segmentation may also be required. In a second phase, the project will focus on two potential areas. The first area is (domain) adaption, i.e., (semi-)automatic adaptation of the visual anomaly detection system to new products and scenarios. In addition, we will focus on improving detection performance at test time (continual learning) by incorporating detected defects or new variations in the normal data distribution (e.g., covariate shifts). A possible second area focuses on enhancing user experience and involves developing anomaly explanation features to make the system more informative and verifiable. |
|
This thesis is conducted and supervised by Prof. Dr. Asja Fischer and Simon Damm from the Chair of Machine Learning (informatik.rub.de/ml) and INI (ini.rub.de) in collaboration with IMS Messsysteme GmbH (ims-gmbh.de) and takes place as a 6-month paid internship at IMS Messsysteme GmbH in 42579 Heiligenhaus.
Please apply to simon.damm@rub.de or pavlos.rath-manakidis@rub.de and include your transcript of records and a CV without a picture.
Prerequisits:
• Completed courses, internships, or projects on machine/deep learning, and/or computer vision.
• Experience with the Python programming language, preferably with machine learning libraries like PyTorch.
• Willingness to work on-site at IMS Messsysteme GmbH in 42579 Heiligenhaus
Main References:
[1] Damm, S., Laszkiewicz, M., Lederer, J. and Fischer, A., 2024. AnomalyDINO: Boosting Patch-based Few-shot Anomaly Detection with DINOv2. arXiv preprint arXiv:2405.14529.
[2] Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T. and Gehler, P., 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 14318-14328).
[3] Ruff, L., Kauffmann, J.R., Vandermeulen, R.A., Montavon, G., Samek, W., Kloft, M., Dietterich, T.G. and Müller, K.R., 2021. A unifying review of deep and shallow anomaly detection. Proceedings of the IEEE, 109(5), pp.756-795.
Additional References:
[4] Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A. and Assran, M., 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193.
[5] Akcay, S., Ameln, D., Vaidya, A., Lakshmanan, B., Ahuja, N. and Genc, U., 2022, October. Anomalib: A deep learning library for anomaly detection. In 2022 IEEE International Conference on Image Processing (ICIP) (pp. 1706-1710). IEEE.