Barlume: Open-Source Document-Image Compression and OCR Degradation — A Reproducible, Multi-Axis Evaluation (Code and Result Tables)
Year: 2026
Type: software
Abstract: Barlume is an open-source reproducibility package for evaluating document-image compression across file size, visual fidelity, OCR degradation, runtime, document category, and language. Version 2.2.0 accompanies the IEEE Access article Open-Source Document-Image Compression and OCR Degradation: A Reproducible, Multi-Axis Evaluation, IEEE Access, vol. 14, pp. 144402-144419, 2026, doi:10.1109/ACCESS.2026.3724788. The package contains benchmark and verification code, page-level derived results for 56,579 compression runs over 1,951 pages, publication tables and figures, sampling manifests, OCR page-eligibility records, and descriptive cross-engine sensitivity outputs. The benchmark covers five configured constructions and 29 settings. OCR-degradation analyses are restricted to 540 of the 1,651 OmniDocBench pages with non-empty normalized ground truth and reference CER ≤ 0.50, yielding 15,660 analyzable rows. DocLayNet contributes 300 layout-diverse pages to file-size, SSIM, and runtime analyses but provides no OCR ground truth. A separate verification implementation reproduces 79 intersecting rows over a union of 1,488 rows (row-level Jaccard index 0.05309139784946237); collapsing the same tail memberships to unique pages yields 17/211 (0.08056872037914692). Source pages are not redistributed. OmniDocBench and DocLayNet must be obtained separately under their respective terms; exact sampled identifiers are included. Original Barlume benchmark code is distributed under the MIT License. The vendored pdf-segmented component retains the GNU Affero General Public License v3.0 or later. Version 2.2.0 provides the final reproducibility materials as a verified, modular download set and adds explicit page-level OCR-eligibility metadata. It accompanies the published IEEE Access article (doi:10.1109/ACCESS.2026.3724788).
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164166316
