Integrating Textual-Based and Visual-Based Features in Poem Detection for Digitized Historical Newspaper Document Images
Journal on Computing and Cultural Heritage · 30.09.2025 · 18 (3) · 1-28
DOI: 10.1145/3746404 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal on Computing and Cultural Heritage, online erschienen am 09.09.2025. Verfasst von Yi Liu und 2 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 87 Literaturangaben.
Abstract (Originalsprache)
Conventionally, automated poem detection using textual features based on semantic and structural characteristics performs well when accurate machine-readable texts are readily available. However, for digitized historical documents such as historical newspapers that are not born digital, performing OCR to obtain machine-readable texts may lead to significant errors because such documents are inherently noisy. On the other hand, while poems can be detected based on their visual cues in layout structures, some visual cues are not unique to poems, leading to incorrect detection. We propose an integrated poem detection approach that combines conventional textual features with visual features. First, we extract 11 numerical representations of textual features using statistical methods and natural language processing techniques based on OCR-generated texts. Second, we extract 20 numerical representations of visual features based on the margin, jaggedness, stanza, and text line length of the text line layouts that are highlighted using a novel projection-based consolidation strategy. Finally, we use a combined vector of the 11 textual and 20 visual features to train a poem detection classifier using an Artificial Neural Network. To evaluate our proposed approach, we use a dataset consisting of nearly 16,000 historical newspaper document images, originally from the Chronicling America repository. The dataset is accessible at https://doi.org/10.13014/K22N506T . Our investigation evaluates the effectiveness of the proposed poem detection approach, showing it outperforms approaches that use textual or visual features alone. Further, to evaluate our proposed projection-based consolidation, we include a smear-based consolidation approach for highlighting text line layouts. Our investigation shows the classifier trained with visual features extracted using the projection-based consolidation outperforms the one using the smear-based consolidation.
Zitieren
APA
Liu, Y., Soh, L., & Lorang, E. (2025). Integrating Textual-Based and Visual-Based Features in Poem Detection for Digitized Historical Newspaper Document Images. Journal on Computing and Cultural Heritage, 18(3), 1-28. https://doi.org/10.1145/3746404
Vancouver
Liu Y, Soh L, Lorang E. Integrating Textual-Based and Visual-Based Features in Poem Detection for Digitized Historical Newspaper Document Images. Journal on Computing and Cultural Heritage. 2025;18(3):1-28. doi:10.1145/3746404
BibTeX
@article{Liu2025,
author = {Liu, Yi and Soh, Leen-Kiat and Lorang, Elizabeth},
title = {Integrating Textual-Based and Visual-Based Features in Poem Detection for Digitized Historical Newspaper Document Images},
journal = {Journal on Computing and Cultural Heritage},
year = {2025},
volume = {18},
number = {3},
pages = {1--28},
doi = {10.1145/3746404}
}
EUVIMED