
Example of text lines segmented in old document
Page segmentation and Layout analysis
Documents are complicated objects. The way text and images are arranged on a page has meaning as much as the words do. The structure of the page needs to be determined, and often the page needs to be broken into its components to be better understood.
Text-line segmentation
We use machine learning models to detect text lines in handwritten historical documents. Our goal is to achieve accurate line segmentation in a zero-shot setting—without fine-tuning on new datasets.
To evaluate this, we compare several machine learning approaches and analyze their performance on line segmentation tasks. We also investigate how few-shot learning can improve results and assess its overall impact.
Form Field Extraction
We use object detection models to extraction information from documents containing forms by locating different field types such as name, address and signature.
Our recent study involved various number of object detection families, and we released a curated version of the CommonForms dataset for this task.
Page Layout Analysis
To insure better model accuracy either during detection and recognition phases, we focus on the page layout and the connection between different page components.
In our work we target page layout analysis to deal with complex page structure to isolate and segment its components such as tables, figures, paragraph, and to distinguish between different textual information such main text, marginalia, title and footer. The datasets we used include modern and historical manuscripts in Latin and Arabic scripts, which even large vision-language models still struggle to understand and process.
Publications
Gayan Pathirage, Stephan Unter, Elisa Barney Smith, "Generalization of Text Line Segmentation for HTR in Historical Documents," ICDAR, 8th International Workshop on Historical Document Imaging and Processing (HIP), September 2026
Iheb Brini, Omar Moured, Hamza Gbada, Elisa Barney, "Mind the Gaps: A Curated Benchmark for Form Field Detection," ICDAR Document Analysis Systems (DAS) Workshop, September 2026
Theses
Jean Robert Twagiramungu, "Separation of Handwritten and Printed Text for Recognition in
Historical Documents at the Region Level Using YOLOv8," Master Programme in Data Science, Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, 2026
Dahlman, Maria, "Text Line Detection in Handwritten Documents using Character Spotting and Hough Transform," Master Programme in Data Science, Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, 2025
Participants
Elisa Barney, Iheb Brini, Gayan Pathirage, Stephan Unter
Contact
Elisa Barney Smith
Updated: