
What happened
Across 15 setups on 11 real Japanese receipts, YomiToku's two-pass read hit 93% exact product names and 100% of amounts at roughly 1.8 seconds per receipt; NDLOCR-Lite and PP-OCRv6 led the commercial-friendly options.
Why it matters
The most accurate option on printed receipts carries a non-commercial license, so anyone planning to embed it in a paid product or workflow must choose a less accurate engine instead.
Summaries like this, every morning.
The test grew out of a personal expense-tracking app: the author wanted store names, dates, items and amounts filled in automatically from photos of receipts, and chose to process images locally rather than send them to an outside service, since receipts carry store, item and even staff names. The machine used is deliberately modest — a Beelink SER10 MAX mini PC with an AMD Ryzen AI 9 HX 470 and integrated Radeon 890M graphics, run through Proxmox with 8 CPU threads and 16GB of memory and no NVIDIA GPU, a setup that departs from the NVIDIA-centric assumptions of most OCR write-ups.
The results split along the nature of the document. Japanese-specialized and conventional engines, which detect where text sits and then read line by line, returned coordinates and confidence scores and almost never invented characters; that let the author re-read only low-confidence lines with a heavier model, which proved both more accurate and much faster than running the heavy model alone. The image-capable VLMs could keep up on short receipts but broke down on the long, curved 450×600-pixel Life supermarket receipt that held 21 of the 43 items, with PaddleOCR-VL looping on "* 198" and DeepSeek-OCR 2 generating inflated amounts up to "49,080". On the two handwritten receipts the ranking roughly reversed, with Qwen3-VL 4B, GLM-OCR and HunyuanOCR taking the top spots while YomiToku fell to the middle.
The author is explicit that this is a small sample — 11 receipts, with half the items concentrated on one — and that a single item can flip the ranking, so the numbers are best read as a tendency. The practical takeaway offered is procedural rather than a single winner: because handwritten receipts have no item column, arithmetic checks cannot catch a misread amount, so reading with two engines and accepting a result automatically only when they agree, with a human reviewing any mismatch, appears the safer path.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
John Deere announced the acquisition of Orchard AI, a San Francisco company whose vehicle-mounted, AI-powered…

LTT Labs tested the RTX Spark-based ASUS ProArt P16 H7607 (64GB) against NVIDIA DGX Spark and AMD Ryzen AI Hal…

Google is upgrading Google Pics — both the standalone app and the feature built into Google Docs/Slides — to N…

The poster for "Artificial" — Luca Guadagnino's film with Andrew Garfield as Sam Altman — struck Instagram use…

JR東日本 raised branded searches 12% using Gemini-generated ad copy and Imagen images; シーズンズ cut CPA 82% with in-…

Kuaishou unveiled Kling Image 4.0 on Nov 18, generating native 8K images, up to 10 coordinated images per prom…
