Work

PosterSentry: a Lightweight Multimodal Classifier for Scientific Poster Quality Assurance in Open Repositories

Year: 2025

Type: article

Abstract: The Posters.science project aims to build an index of findability, accessibility, interoperability, and reusability (FAIR) of posters by extracting machine-actionable metadata from PDFs stored in open access repositories. During harvesting of poster records from Zenodo and Figshare, we found that many were mislabeled, containing multi-page papers, proceedings, abstract books, or presentations. Label noise at this scale severely limits the viability of automated FAIR pipelines. To mitigate this, we developed PosterSentry, a lightweight, open-source multimodal classifier that identifies whether a PDF is a true scientific poster. We fused three complementary feature streams in PosterSentry: 1) text embeddings via model2vec (512 dimensions), 2) visual features including color statistics, edge density, and spatial complexity (15 dimensions), and 3) structural PDF layout features such as page geometry, font diversity, and text density (15 dimensions). We concatenated these streams into a 542-dimensional vector and used logistic regression to classify them. PosterSentry was trained on 3,606 heuristically selected real-world examples with no synthetic augmentation. It achieved 87.3% accuracy, runs on CPU in under one second per file, and requires only a 10kb model head, which is orders of magnitude smaller than other LLM-based alternatives. PosterSentry is publicly available under an MIT license on Hugging Face alongside its training data, and is currently deployed in the Posters.science quality control pipeline. In this presentation, we will present the development of PosterSentry, results across 30,205 poster PDFs from open repositories, and discuss future work planned in collaboration with the Collaborative Metadata (COMET) initiative.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.21842321

Authors: ONeill, James, Sarin, Parth, Buttrick, Adam, Patel, Bhavesh

Institutions: California Medical Innovations Institute, Stanford University, California Digital Library

Language: English

Topic: Handwritten Text Recognition Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funder: The Navigation Fund

Award: G5647235611