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

Sergei Solovev

2026-02-26 · Preprint, Figshare · DOI: 10.6084/m9.figshare.31430086

Download PDF View on Figshare

Abstract

EasyOCR + heuristic rules vs. Donut end-to-end Transformer on SROIE 2019. Error taxonomy under image degradation typical of messenger-grade distortions (compression, rotation, blur).

Key questions

Q. OCR pipeline vs end-to-end Transformer (Donut) for receipt information extraction — which is more robust?
The paper compares EasyOCR with heuristic rules against the Donut end-to-end Transformer on SROIE 2019 and builds an error taxonomy under image degradation typical of messenger-grade distortion — compression, rotation and blur. See the paper for the per-condition breakdown.

Q. Why test receipt OCR under image degradation?
Real-world receipts often arrive as compressed, rotated or blurred photos; the study characterizes how OCR-based and end-to-end pipelines fail under exactly those messenger-grade distortions.

Keywords: OCR; Donut transformer; document understanding; SROIE 2019; receipt extraction; image degradation; error taxonomy