Index

Evaluate methods to improve few-shot classification on CXRs. This is a rarely studied area for medical domains.

Evaluating State of the Art Scene Text Recognition Methods for Recognizing Text on Tombstones

Investigating Target Class Influence on Neural Network Compressibility for Edge Deployment in Avian Species Identification

Cross-Process Anomaly Detection in Multivariate Time Series for Minimising Quality Drift in Electric Powertrain Production: A Predictive Quality Approach

Manufacturing processes inherently display variability, no matter how precise and delicate the
process may be. Correctly identifying these variations in production helps enhance product quality,
promote sustainability, and improve employee morale by creating a more consistent and predictable
work environment. This Master’s thesis will use methods in Machine Learning (ML) and Statistics
to trace quality drifts and anomalies in production, potentially caused by complex dynamics of
mechanical and chemical processes in rotor production. This thesis aims to make a meaningful
contribution towards the ultimate goal of Zero Scrap in manufacturing processes. There is minimal
research on applying ML to press-in curve (force/ displacement) anomaly detection. State-of-theart
methods for detecting the generation of quality variations and defects consider only production
processes at a single stage at a time. However, there are research initiatives such as Project-
MUPROD [2014], Meiners et al. [2021], Kornas et al. [2019] to prevent the transmission of defects
between processes at a multistage level. Yet they suffer from not attaining precision and cannot
model temporal dependencies across series production processes. This thesis divides its objectives
into the following to address these shortfalls:

• To conduct a systematic literature review, gather relevant data from distributed sources, and
develop an analysis pipeline.
• To sequentially analyse regularly and irregularly sampled multivariate time series data at
multiple stages and perform anomaly detection with each stage as a quality checkpoint and
End-of-Line(EOL) as the final quality determinant. To compare a few methods.
• To identify root causes of a failure and recommend optimal process parameters/ reach new
process thresholds. Determine whether quality variations stem from a single stage or a
combination of stages in the production process.
• To model cross-process interactions with memory among multiple press-in processes using
Deep Learning architectures such as Recurrent Neural Networks or Transformers.

References
T. Kornas, R. Daub, M. Z. Karamat, S. Thiede, and C. Herrmann. Data-and expert-driven
analysis of cause-effect relationships in the production of lithium-ion batteries. In 2019 IEEE
15th International Conference on Automation Science and Engineering (CASE), pages 380–385,
2019. doi: 10.1109/COASE.2019.8843185.
M. Meiners, A. Mayr, and J. Franke. Process curve analysis with machine learning on the example
of screw fastening and press-in processes. Procedia CIRP, 97:166–171, 02 2021. doi: 10.1016/j.
procir.2020.05.220.
Project-MUPROD. Innovative proactive quality control system for in-process multi-stage defect reduction.
CORDIS Europa, Oct. 2014. URL https://cordis.europa.eu/project/id/285075.

RetNetHTR : Leveraging Retentive Networks for Efficient and Accurate Handwritten Text Recognition

Semi-Supervised Learning for Glacier Front Delineation

The use of synthetic aperture radar (SAR) imagery allows for all year round monitoring of glacier
movements, regardless of weather influences. This results in huge amounts of data, making human
evaluation of every image infeasible. Advances in the field of Deep Learning create new ways for
automatic image segmentation using a variety of models. Using the CaFFe benchmark dataset [3]
allows for a proper comparison of different model architectures. The problem of unlabeled data
still persist, with the CaFFe training dataset only consisting of only five different glaciers with
some of them being underrepresented.
In order to take advantage of unlabeled data, this thesis will apply semi-supervised learning techniques.
Semi-supervised learning is, as the name suggests, a hybrid learning scheme of supervised
and unsupervised models, using both labeled and unlabeled SAR images. After training the model,
the probability for each pixel in the image to belong to one of four classes (glacier, ocean, rock
outcrop and areas with no information available) is calculated, resulting in a zone prediction. The
front of the glacier can then be calculated afterwards during the post-processing. Two of the most
famous self-supervised learning schemes are iBot [6] and DinoV2 [5], which will both be analyzed
and evaluated. Both frameworks will have to be modified according to the guidelines laid out by
Gourmelon et al. [2], in order to properly compare them with each other and with other models
trained on the CaFFe dataset.
The backbone used in both cases is the HookFormer, a model based on the Swin Transformer [4]
that has shown better performance when used for glacier front delineation [1]. It employs two
Transformer models with a cross-resolution interaction between them, using images of different
resolution from the same area.

References
[1] Nora Gourmelon et al. Deep learning has yet to match human performance in delineating
glacier calving fronts. In Preparation.
[2] Nora Gourmelon, Thomas Seehaus, Matthias Braun, Andreas Maier, and Vincent Christlein.
Calving fronts and where to find them: a benchmark dataset and methodology for automatic
glacier calving front extraction from synthetic aperture radar imagery. Earth System Science
Data, pages 4287–4313, 2022.
[3] Nora Gourmelon, Thorsten Seehaus, Julian Klink, Matthias Braun, Andreas Maier, and Vincent
Christlein. Caffe – a benchmark dataset for glacier calving front extraction from synthetic
aperture radar imagery. In IGARSS 2023 – 2023 IEEE International Geoscience and Remote
Sensing Symposium, pages 896–898, 2023.
[4] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining
Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021
IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992–10002. IEEE
Computer Society, 2021.
[5] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov,
Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran,
Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra,
Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick
Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features
without supervision. arXiv preprint arXiv:2304.07193v2, 2024.
[6] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong.
ibot: Image bert pre-training with online tokenizer. International Conference on Learning
Representations (ICLR), 2022.