OCR-Based vs. End-to-End Transformer Pipelines for Receipt Information Extraction: A Comparative Study on SROIE 2019

We present a systematic comparison of two fundamentally different paradigms for key information extraction (KIE) from scanned receipts: (1) a cascaded OCR pipeline combining EasyOCR text recognition with rule-based field extraction, and (2) an end-to-end vision-language transformer (Donut) that maps raw pixels directly to structured output without intermediate OCR. Both approaches are evaluated on the ICDAR 2019 SROIE benchmark for extracting three target fields — vendor, date, and total — under both clean and messenger-grade corrupted image conditions. Experiments on 80 validation receipts reveal that no single paradigm dominates: the fine-tuned Donut model achieves the highest overall micro-F1 (0.75 vs. 0.63), yet the OCR pipeline outperforms it on date extraction (F1 = 0.78 vs. 0.63) owing to the effectiveness of regex patterns on closed-vocabulary fields. A 12-category error taxonomy shows that the two pipelines fail on largely non-overlapping subsets of documents (Pearson r = 0.30), indicating strong potential for ensemble strategies. The end-to-end model also demonstrates superior robustness to image degradation (ΔF1 = −0.12 vs. −0.17), validating theoretical predictions about cascading error propagation in multi-stage systems. All code, data, and trained models are publicly available at https://github.com/SergeySolovyev/Invoice-DocAI We release a reproducible benchmark pipeline, code, and trained checkpoints, and discuss practical guidelines for choosing between OCR-based and OCR-free approaches under real-world image corruption. --- Cite as: Solovev, S. (2026). OCR-Based vs. End-to-End Transformer Pipelines for Receipt Information Extraction: A Comparative Study on SROIE 2019 (Preprint). Figshare. https://doi.org/10.6084/m9.figshare.31430086