Bachelor and Master Theses

Current machine learning pipelines specially working with image and computer vision heavily depend on large quantities of labelled dataset. Recent advances in self-supervised learning (SSL) and active learning (AL) can help to reduce the need for manual data labelling. SSL leverages inherent structures in unlabelled data to learn useful representations without human supervision, while AL strategically selects the most informative samples for manual annotation. Combining these two paradigms can potentially minimize human involvement in labelling while maintaining or even improving model performance.

The objective of this thesis is to develop an automated image labelling tool utilizing self-supervised learning and active learning approaches.

Expected Outcomes

- Significant reduction in manual image annotation/labelling effort compared to traditional methods.

- Improved model accuracy with fewer labelled samples

• Python GUI programming and PyTorch, Scikit-learn libraries
• Knowledge on image processing, convolutional neural networks and CNN architectures  
• Working with benchmark image datasets, e.g COCO, CIFAR, ImageNet
• Data preprocessing and augmentation, data handling and visualization