Quantifying the Effects of Image Degradation on LVLM Benchmark Results Systematically

Rupert Urbanski, Ralf Peters

2025

Abstract

Degraded image quality, along with the underlying issues of text-to-text neural networks, can compromise the performance of LVLMs. This paper quantifies the impacts of blurry, noisy and warped images and evaluates the robustness of LVLMs towards the common forms of image degradation in real-world applications utilising a specifically developed benchmark dataset comprising 15840 systematically degraded text images, which were hand-crafted based on standardised university admission exams.

Paper Citation

in Harvard Style

Urbanski R. and Peters R. (2025). Quantifying the Effects of Image Degradation on LVLM Benchmark Results Systematically. In Proceedings of the 14th International Conference on Data Science, Technology and Applications - Volume 1: DATA; ISBN 978-989-758-758-0, SciTePress, pages 355-362. DOI: 10.5220/0013462800003967

in Bibtex Style

@conference{data25,
author={Rupert Urbanski and Ralf Peters},
title={Quantifying the Effects of Image Degradation on LVLM Benchmark Results Systematically},
booktitle={Proceedings of the 14th International Conference on Data Science, Technology and Applications - Volume 1: DATA},
year={2025},
pages={355-362},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013462800003967},
isbn={978-989-758-758-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 14th International Conference on Data Science, Technology and Applications - Volume 1: DATA
TI - Quantifying the Effects of Image Degradation on LVLM Benchmark Results Systematically
SN - 978-989-758-758-0
AU - Urbanski R.
AU - Peters R.
PY - 2025
SP - 355
EP - 362
DO - 10.5220/0013462800003967
PB - SciTePress