Deep learning techniques for image segmentation and anomaly detection in low data regimes

This thesis focuses on the problem of image segmentation in low-data settings. In particular, the two specific problems that are tackled in the present work are the ones revolving around anomaly segmentation for industrial quality control and doc- ument layout segmentation of ancient manuscripts. For the first problem, two novel attention-based are proposed one based on the popular U-Net architecture and the second one on the more recent Vision Transformer which has been enhanced for the task at hand with a masking module and a multi-resolution self-attention compo- nent. As for the document layout analysis, we introduce a few-shot segmentation framework based on the combination of DeepLabV3+, a robust deep learning ar- chitecture for semantic segmentation, with a traditional computer vision algorithm for image binarization while at the same time relying on a novel instance generation strategy that allows to leverage the small amount of data available fully. Further- more, we provide an analysis of the effects of transfer learning in this domain-specific context, showing the drawbacks of pre-training on large general-purpose datasets compared to smaller domain-specific ones. For each of the proposed approaches, we provide the experimental results obtained on popular publicly available datasets for the corresponding task.