Receipt OCR Accuracy: Thermal Paper, Crumpled, Faded,
and Multi-Currency Benchmarks (2026)

Breakdown by Receipt Condition: The Degradation Ladder

Receipt OCR accuracy is governed less by the engine than by the physical state of the paper and how it was captured. The same receipt moves down the ladder — fresh and flat, faded, crumpled, phone-photographed — and each step costs 5 to 30 accuracy points.

Receipts differ from invoices in a way that matters more than layout: the medium itself is unstable. Most receipts are printed on thermal paper, a heat-sensitive coating that fades progressively with exposure to light, heat, and time — a receipt legible the day of purchase can have partially unreadable tax amounts, dates, and line items weeks later, and no OCR engine can recover text that has physically vanished from the paper. The SROIE organizers explicitly list "poor paper quality, poor ink and printing quality, low resolution scanner and scanning distortion, folded invoices" among the benchmark's core challenges (Huang et al., 2019). The chart below plots the published accuracy band for each rung of the ladder; the metric used by each study is labeled on the row, because character-, word-, and field-level numbers are not interchangeable. The SROIE figures are F1 scores — the harmonic mean of precision and recall (Hmean), which requires both that most detected words be correct and that most ground-truth words be found.

Sources: IJSAT (2026) — EasyOCR at 300 DPI on printed documents, character accuracy (flat 94.7%, crumpled 86.9%; the same study measured dirty 80.9% and wet 72.7%); KORIE (2026) — word accuracy on 748 degraded Korean thermal receipts, 17,587 crops, open-source engines (Tesseract 64.7%, EasyOCR 68.6%, PaddleOCR 73.3%); RMIT Textract study (2025) — exact-match merchant-name field accuracy on 118 real receipts (65% phone photos, 50% folded/crumpled, 44% skewed). Character-level figures are not directly comparable to word- or field-level figures — see Why the Metrics Matter.

Sources listed in table above. The KORIE rows are the only published benchmark that preserves real thermal-paper artifacts (ink fading, print-head banding, friction wear, creases) rather than synthetic degradation; its 300-DPI flatbed scans mean even the "scanned" best case embeds the paper's own decay. The SROIE ceiling row is 2019-era technology — modern LLM engines on the same dataset reach 84.3–93.0% overall field accuracy (Doubleword, 2026; see engine-tier table).

Breakdown by Field Type: Total Is Easy, Merchant Is Hard

Field type is the second largest driver, and it is remarkably consistent across every study that reports it: numeric summary fields — especially the total — extract far more reliably than text-identity fields such as merchant or vendor name. The chart below plots five modern LLM engines on the full 626-receipt SROIE test set, all measured with the same field-level exact-match protocol (Doubleword, 2026).

Sources: Doubleword (2026) — 626 SROIE test receipts, field-level exact match, 5 engines (Qwen3-VL-235B/30B, GPT-5-mini/5.2/5-nano). D-grade vendor benchmark with fully disclosed methodology; corroborated in direction by ReceiptSense (2024), which found numeric #Units fields reach 93.42 F1 while text Brand fields fall to 62.30 F1 on Arabic–English receipts.

The pattern survives across radically different conditions and languages. The RMIT Textract study — real phone-captured receipts, not a clean corpus — found the total was "consistently detected" while merchant names misfired (misclassified nearby text, or "Unknown" when absent), matching the Doubleword ranking exactly (2025). The practical implication is that a "receipt OCR accuracy" headline averages the easiest and hardest fields into a single number — and expense workflows that only need the total are quoting a fundamentally easier task than workflows that need merchant plus line items.

Breakdown by Engine Tier

Engine choice matters less than condition and field type, but it sets the ceiling each condition can reach. The table below separates four tiers with incompatible measurement contexts: research models fine-tuned on receipt corpora (field-level F1 on the CORD/SROIE benchmarks), general-purpose LLMs without fine-tuning, cloud APIs on real-world documents, and legacy/traditional engines.

Sources listed in table above. CORD and SROIE figures are field-level F1 on clean benchmark corpora where the model was fine-tuned on the task — they are the optimistic ceiling, not the real-world expectation. The real-world rows (Google forum, RMIT Textract) measure different document populations and are not directly comparable; both are labeled with their conditions. The DATABASICS figure is a single vendor's tier claim and appears only as a floor reference.

Multi-Currency and Multi-Language Receipts

Global expense workflows add a third dimension: every benchmark dataset on this page is monolingual or bilingual, and the only multilingual study published with per-field numbers shows the same total-vs-text gap plus an extra script penalty. On 20,000 Arabic–English receipts, general LLMs without fine-tuning reached only 32.07–42.98 F1 zero-shot, rising to 60.60–80.26 F1 with three examples per field — while numeric #Units fields hit 93.42 F1 against 62.30 F1 for text Brand fields (ReceiptSense, 2024). The RMIT study observed the same script effect in production: receipts with non-English product names were "partially parsed," with item data silently dropped (2025).

Sources listed in table above. No peer-reviewed study isolates currency effect (symbol, decimal convention, three-letter codes) while holding script constant — the "multi-currency" gap in real workflows is driven primarily by script and layout variation, with the numeric value itself extracting well once recognized. This is flagged as a data gap in Limitations rather than filled with vendor numbers.

Why the Metrics Matter: Character, Word, and Field Are Three Different Answers

The wide bands on this page are not measurement noise — they are three different questions wearing the same word "accuracy." A receipt can score 98% character accuracy, ~85% word accuracy, and ~75% field accuracy at the same time, and all three numbers are true.

Character accuracy counts individual characters: a 1% error rate means one wrong character per hundred. Word accuracy fails a whole word on any single misread character — measured as Word Error Rate (WER), which is why KORIE's best engine shows only 15.84% character error (CER) but a 26.73% WER, nearly double the error at word level (KORIE, 2026). Field accuracy fails an entire field — merchant name, date, total — on any wrong character or token in it, which is why the SROIE organizers concluded that "receipt OCR has much higher accuracy requirements than the general OCR tasks" and why field-level F1 is the metric expense workflows actually consume (Huang et al., 2019). Every figure on this page is labeled with its level; blending them into one "receipt OCR accuracy" number is exactly how the 99% claim is manufactured.

How to Use This Data

You don't need a pilot project to estimate what receipt OCR will deliver on your expense flow. A four-step back-of-the-envelope calculation using the ranges above turns "vendors say 99%" into a defensible expected error count — and it usually reveals that the conditions of your receipts, not the engine, set the outcome.

These are rough-estimate formulas, not a substitute for benchmarking your own documents. Every published range on this page was measured on someone else's corpus — the only accuracy number that applies to your receipt flow is the one you measure on it. The value of this page is setting the expectation before you measure, and knowing which metric to ask for.

Frequently Asked Questions

How accurate is receipt OCR?

The honest answer is a ladder, not a number: 90%+ word-level on clean scanned receipts (best 2019 SROIE team, Huang et al.), 55–73% word accuracy on real faded thermal receipts (KORIE 2026), and 66.7–68.7% field accuracy on merchant names from real phone-captured receipts (RMIT Textract study 2025). Vendor claims of 97–99% describe fresh, flat, well-lit captures of easy fields — the top of the ladder, not the middle.

Can OCR read faded thermal paper receipts?

Partially — and the loss is physical: faded thermal text is permanently gone from the paper, so no engine can recover it. On Korean thermal receipts that retained real fading and creasing artifacts, open-source engines scored only 55–73% word accuracy (KORIE 2026) — the best (PaddleOCR, 73.3%) still fails roughly one word in four. The only reliable fix is capturing receipts before they fade.

Does crumpling a receipt affect OCR accuracy?

Yes — measurably. In a controlled 300-DPI test, crumpling dropped character accuracy from 94.7% to 86.9%, about 8 percentage points; dirt and water damage drop it further to 80.9% and 72.7% (IJSAT 2026). The RMIT study's real-world receipts were ~50% folded or crumpled, consistent with its ~67% merchant-field result.

Which receipt field is hardest for OCR to extract?

The merchant (vendor) name is consistently the hardest, at 73.1–87.7% field accuracy across five LLM engines, while the total is the easiest at 94.9–98.9% and the date sits between at 84.3–92.5% (Doubleword 2026, 626 SROIE receipts). The same ranking appears on real phone-captured receipts (RMIT 2025) and on Arabic–English receipts (ReceiptSense 2024: numeric fields 93.4 vs text fields 62.3 F1).

Is receipt OCR less accurate than invoice OCR?

Yes, in practice. Invoices arrive as controlled PDFs or flatbed scans with standardized layouts; receipts are thermal paper, phone-photographed, crumpled, and layout-free. The gap shows up in comparable cloud engines: on invoices, field accuracy runs 78–98% by engine (BusinessWareTech 2025), while the same engine class on real phone-captured receipts scored 66.7–68.7% for the merchant field (RMIT Textract study 2025).

How much does phone capture hurt receipt OCR accuracy vs scanning?

There is no peer-reviewed controlled comparison on identical receipts — a genuine data gap (see Limitations). The evidence that exists: the RMIT study's receipts were ~65% phone photos with 44% skewed and 12.7% blurred, yielding 66.7–68.7% merchant-field accuracy, while the SROIE and KORIE benchmarks that score higher used flatbed scans. Directionally, phone capture adds skew, shadows, and blur that each push accuracy down the ladder, but the exact penalty is unmeasured in the published literature.

What is SROIE and why does it matter?

SROIE (Scanned Receipt OCR and Information Extraction) is the ICDAR 2019 competition that created the first standardized receipt OCR benchmark: 1,000 real scanned receipts with text-localization, OCR, and key-field (company, address, date, total) tasks (Huang et al., 2019). It matters because it is the only receipt benchmark with published cross-team results: even the best 2019 method fell short of the 99% accuracy that receipt applications demand. Modern LLMs on the same dataset score 84.3–93.0% overall (Doubleword 2026).

Methodology & Sources

How This Page Was Built

This page aggregates data from 14 independent studies spanning 2019–2026. Sources were selected based on three criteria: (1) publicly documented methodology with stated sample size, (2) data collected within the last 3 years where available, with older sources labeled by year in-text, and (3) relevance to receipt-specific OCR and extraction rather than generic document claims. Where multiple sources report the same dimension, conservative ranges (lowest–highest reported values) are used — we prefer to under-claim. Where sources disagree significantly, the disagreement is explained by metric level, document condition, or corpus rather than averaged away.

The metric distinction is the spine of this page: character-level, word-level, and field-level figures are never blended into a single number, and each chart and table row labels its metric. Vendor self-reported figures (DATABASICS ~64% tier floor) are labeled as D-grade and used only as corroborated references. The SROIE benchmark results are 2019-era technology and are labeled by year; the same dataset's modern LLM results (Doubleword 2026) are shown alongside for currency.

Source List

1. Huang, Z. et al. — "ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction" (2019). Peer-reviewed competition report; 1,000 scanned receipts, 29/24/18 valid submissions across three tasks. Provides the 90.49% best Task-3 F1, the "only one of 18 teams above 90%"/"more than half below 80%" findings, and the Task-2 "even the best method cannot reach the 99% requirement" statement.
2. Park, S. et al. — "CORD: A Consolidated Receipt Dataset for Post-OCR Parsing" (2019). Academic dataset paper; 11,000+ Indonesian receipts, 30 fields under 4 superclasses; public subset 800/100/100. Provides the corpus definition and field-level F1 evaluation protocol used by the research-model rows.
3. Xu, Y. et al. — "LayoutLMv2" (ACL 2021). Peer-reviewed paper. Provides CORD field F1 96.01% (426M parameters) and SROIE field F1 97.81%.
4. Kim, G. et al. — "Donut: OCR-free Document Understanding Transformer" (ECCV 2022). Peer-reviewed paper and official model repository. Provides CORD field F1 91.3% (donut-base-finetuned-cord-v2).
5. "KORIE: A Multi-Task Benchmark for Detection, OCR, and Information Extraction on Korean Retail Receipts" — MDPI Mathematics (2026). Peer-reviewed journal; 748 Korean thermal receipts with real fading/banding/crease artifacts; 17,587 OCR crops. Provides the only real-degradation word figures: PaddleOCR WER 26.73% (73.3% word accuracy), EasyOCR 31.43%, Tesseract 35.26%, BiGRU 44.47%.
6. "Towards Analysing Invoices and Receipts with Amazon Textract" (RMIT, 2025). Academic case study (preprint); 118 Australian receipts, ~65% phone-photographed, ~50% folded/crumpled, 44% skewed. Provides merchant-name exact-match 68.7% (text layouts) / 66.7% (logo layouts), the "totals consistently detected" finding, and the language-limitation observation.
7. Abdallah, A. et al. — "ReceiptSense: Beyond Traditional OCR" (2024). Academic preprint; 20,000 Arabic–English receipts, 30,000 OCR images. Provides zero-shot F1 32.07–42.98%, three-shot F1 60.60–80.26%, and the numeric-vs-text field gap (93.42 vs 62.30).
8. "Optical Character Recognition Accuracy on Degraded Documents" — IJSAT 17(2) (2026). Peer-reviewed; EasyOCR at 300 DPI on four physical conditions. Provides the flat 94.7% / crumpled 86.9% / dirty 80.9% / wet 72.7% character-accuracy ladder used for the physical-damage rows.
9. "Deep Learning based Key Information Extraction from Business Documents: Systematic Literature Review" (2024). Academic systematic review. Provides the average field F1 across published papers: CORD 0.94, SROIE 0.95.
10. "Large language model driven transferable key information extraction mechanism for nonstandardized tables" (2025). Peer-reviewed; LLM-TKIE, no fine-tuning, 3-shot. Provides CORD F1 80.9 / SROIE F1 83.9 as the zero-setup LLM baseline.
11. Doubleword — "Structured Data Extraction" SROIE benchmark (2026). Vendor benchmark with fully disclosed methodology; full 626-receipt SROIE test set, field-level exact match, 5 LLM engines. Provides overall 84.3–93.0% and the per-field bands (total 94.9–98.9%, date 84.3–92.5%, merchant 73.1–87.7%).
12. Google Developer Forum — "Low F1 Score with Document AI on Diverse Invoice and Expense Layouts" (2024). Anecdotal user measurement on Google's official forum; ~1,500 diverse documents. Provides the F1 0.75–0.85 cloud-API real-world band.
13. BusinessWareTech — "AWS Textract vs Google, Azure, and GPT-4o: Invoice Extraction Benchmark" (2025). Independent third-party benchmark (C-grade, transparent method, real invoices). Provides the invoice-context engine ladder: GPT-4o+OCR 98%, Azure 93%, Google 82%, Textract 78% — the easier upper bound for receipts.
14. DATABASICS — "Receipt Capture & OCR for Expense Reports" white paper. Vendor white paper (D-grade); used only as a floor reference. Provides the ~64% traditional-OCR / 85–95% AI-enhanced / 97–99% LLM tier framing, with the caveat that accuracy varies with image quality, layout, and handwriting.

Limitations

• Geographic coverage: Sources cluster in Asia-Pacific and the Middle East (Malaysia/Singapore, Indonesia, Korea, Egypt/Arabic, Australia). No receipt-specific academic benchmark was found for North America, Latin America, or Africa — US/UK receipt OCR accuracy is currently only documented via vendor claims and anecdotal forum reports, not peer-reviewed studies.
• Methodology constraints: Metric levels (character / word / field) differ across studies and are labeled per row rather than blended; several key figures rely on single studies (KORIE thermal degradation; RMIT Textract merchant field; ReceiptSense multilingual), and the Doubleword and Google-Forum figures are D/C-grade (vendor or anecdotal) used with their grade disclosed. The IJSAT crumpled/faded figures come from printed documents, not receipts — receipts on thermal paper are expected to degrade worse, but that is an inference, stated as such.
• Temporal gaps: The SROIE competition data is from 2019; its absolute figures reflect 2019-era methods and are labeled by year, with modern LLM results on the same dataset (Doubleword 2026) provided alongside. No newer standardized receipt competition with published cross-team results was found after SROIE 2019.
• What we could not find: (1) any peer-reviewed controlled study comparing phone capture vs flatbed scan on identical receipts — the single most cited variable in vendor marketing; (2) a receipt-specific line-item extraction benchmark (the only line-item figures are invoice-context, 40–82% by engine, BusinessWareTech 2025); (3) any study isolating currency effect (symbol, decimal convention) while holding script constant; (4) benchmarks for handwritten receipt additions (tips, notes, corrections); (5) public benchmarks for e-receipts/screenshots vs photographed paper. These gaps are stated rather than filled with unverifiable vendor numbers.

Related references: OCR Accuracy by Document Type · Handwriting Recognition Accuracy · Document Processing Cost Breakdown · Manual Data Entry Error Rates

Related reading: The Cost of Manual Handwritten Receipt Data Entry · How AI Handwriting Extraction Accuracy Keeps Improving