Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/A0DA26EE-1CBA-4532-ACE2-2C8F87C703FD | ||
| Year | 2025 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Charilaos Kapelonis, "Automated mosaic tesserae segmentation using artificial intelligence techniques", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2025 https://doi.org/10.26233/heallink.tuc.104765 | ||
| Appears in Collections |
Art is widely recognized as a reflection of civilization and mosaics represent an important part of cultural heritage. Mosaics are an old art form where a surface is decorated by attaching small pieces of material, called tesserae, to a base with adhesive. Due to their age and fragility, they are prone to damage, highlighting the need for digital preservation.This thesis addresses the problem of digitalizing mosaics by segmenting the tesserae, producing a binary image in which they are clearly separated from the background. This task falls under the field of Image Segmentation, one of the primary tasks in Computer Vision. We aim to create a method that achieves it automatically and with high accuracy.Although classical approaches on Image Segmentation yield decent results, we employ Deep Learning models and, more specifically, our work utilizes the latest Segment Anything Model 2 (SAM 2) by META AI, a foundation model that outperforms most conventional segmentation models. Our work involves manually annotating a set of mosaic images to create a dataset which we utilize both to fine-tune the model and to eventually evaluate its performance.The quantitative results, including the evaluation metrics, show a significant improvement compared to the baseline SAM 2 model. The baseline SAM 2 model - evaluated on the large checkpoint - yields 89.00% IoU and 92.12% Recall, while our model achieves 91.02% IoU and 95.89% Recall on the same dataset. Furthermore, we show that our approach can yield even more impressive results if trained without computational and time limitations.This task belongs to the field of Computer Vision problems, hence the qualitative results, such as visual comparative representations, are equally important. These visualizations further confirm that the fine-tuned model we employ brings us closer to an effective solution to this problem.Keywords: Mosaics, Deep Learning, Computer Vision, Image Segmentation, SAM 2, Transfer Learning, Digital Preservation
Copyright © DIAS 2013
