Open-Source Document-Image Compression and OCR Degradation: A Reproducible, Multi-Axis Evaluation

Romero B.V.R.; Lapa B.F.C. · IEEE Access (2026) · DOI: 10.1109/access.2026.3724788

Resumen

Document-image compression must reduce storage and transfer costs without damaging visual inspection or optical character recognition (OCR) reuse. Five configured free and open-source software constructions (four toolchains, 29 settings) were applied to 1,951 pages from OmniDocBench and DocLayNet, yielding 56,579 page-setting compression records; 56,514 successfully rendered rows informed visual analyses. DocLayNet contributes 300 pages but has no OCR ground truth. Signed character-error-rate change against ground truth (CERΔ) was therefore analyzed only for 540 OmniDocBench pages with informative ground truth and reference CER ≤ 0.50 (15,660 page-setting rows; 27.7% of the full corpus). Paired per-page Friedman tests were significant for size, SSIM, and CERΔ (all p < 0.001 ). The custom P2 DjVu construction at fg200-bg25 gave the smallest median file size (19.2 KB/page), while the single-layer JPEG q95 baseline gave the highest median structural similarity (SSIM, 0.9712). Within the common OCR-analyzable universe, the lowest-SSIM and highest-CERΔ 5% average-rank tails had a row-level Jaccard index of 0.053. CERΔ exceeded the author-defined 0.02 threshold in 34.9% of analyzable rows (page-clustered 95% CI 32.7-37.1%). Median-zero settings were not uniformly direction-symmetric: all five P4 settings showed significant positive sign imbalance, whereas P3 q65-q95 did not. At q75, lossy JBIG2 foreground coding increased OCR degradation (marginal Cliff's δ=0.386 ; median paired increase 0.0369). These results support setting-level, category-stratified, page-aware evaluation and local OCR quality assurance; they do not establish performance on OCR-excluded pages or institutional archival collections. Supporting tables and scripts are provided in the Barlume reproducibility package.

Instituciones: Universidad Católica Sedes Sapientiae; Universidad Nacional del Centro del Perú