Prins, Daniël (2026) Looks Can Be Deceiving: Improving Document Layout Analysis of Historical Documents via Inter-Paragraph Textual Continuity. Bachelor's Thesis, Artificial Intelligence.
|
Text
BachelorThesisAIDanielPrins.pdf Download (2MB) | Preview |
Abstract
Understanding historical documents is challenging due to the nature of their physical structure: hard-to-read handwriting, ink bleed-through, and faint text. The first step in this process is Document Layout Analysis. This step is complicated by the fact that different paragraph regions look visually similar but are semantically distinct and belong to different categories. This thesis investigates whether the physical structure of a book can be used to improve predictions by handling cross-page continuity. The proposed pipeline consists of three separate stages, each building on the previous one: a YOLO-based oriented bounding box detector, with a Llama 3.1 language model for semantic reclassification of detected regions, and a fine-tuned GysBERT cross-encoder which predicts whether adjacent paragraphs form a single continuous text, in which case they belong to the same class. Evaluated on document 0014 of Collectie Overijssel, the visual baseline yields a low macro F1 score of 0.141 due to severe class imbalance. Integrating semantic Llama-based reclassification substantially increases the macro F1 score to 0.494, confirming that textual content is the primary driver of classification performance. The addition of GysBERT cross-page continuity predictions yields an additional incremental improvement, raising the final macro F1 score to 0.531. These results demonstrate that combining multi-modal text context with cross-page linguistic continuity offers a robust architectural pattern for making handwritten historical administrative archives searchable and accessible.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Dhali, M.A. and Schuijlenburg, K. van |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 05 Aug 2026 06:43 |
| Last Modified: | 05 Aug 2026 06:43 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37481 |
Actions (login required)
 |
View Item |