Umi-OCR is a free, open-source desktop application for recognizing text locally on Windows and Linux. It supports screenshot capture, batch images, document/PDF processing, searchable double-layer PDFs, QR/barcodes and optional engine plugins such as formula recognition. The official repository uses the MIT license and distributes portable builds with checksums.
Its strongest advantage is operational privacy: a configured local engine can process documents without sending their pixels to a hosted OCR service. That does not make every surrounding action offline. Downloads, update checks, third-party plugins, user-configured HTTP interfaces, synced folders or downstream translation can create network or data-sharing paths. Verify the exact setup with a network monitor when confidentiality matters.
Which workflow does each feature solve?
| Mode | Best use | Output/evidence | Main limitation |
|---|---|---|---|
| Screenshot OCR | Copying text from an app or image | Immediate recognized text | Small UI text and scaling artifacts |
| Batch image OCR | Scans, receipts and page folders | Per-file text with repeatable settings | Reading order across complex layouts |
| Document/PDF | Scanned archives and books | Text export or searchable double-layer PDF | Tables, footnotes and hidden-text alignment |
| QR/barcode | Decode machine-readable symbols | Payload shown for review | Payload can be a malicious URL |
| Formula plugin | Convert isolated math regions | LaTeX-like representation | Symbols and structure need visual verification |
| HTTP interface | Local automation/integration | Machine-readable OCR result | May expose documents if bound beyond localhost |
The OCR pipeline and its independent errors
source page / screen
|
crop + rotate + deskew
|
v
text detection boxes
|
v
line recognition
|
v
layout sorting / ignore regions
|
.-----+-----------.
v v
plain text searchable PDF
| |
human check: names, numbers, negation, reading order
Detection can miss a region even when recognition is accurate. Recognition can misread a character inside a correct box. Layout sorting can arrange correct lines in the wrong order. A searchable PDF can look visually identical while its invisible text layer contains errors. Inspect each layer separately instead of treating readable page appearance as OCR correctness.
Paddle versus Rapid builds
Official release notes describe the Paddle package as faster/higher-resource and suitable for stronger machines, while the Rapid build uses less memory and has broader CPU compatibility but may be slower. Older release text specifically warned that Paddle could fail on some Pentium, Celeron and Atom CPUs. Current v2.1.5 artifacts list a Windows Rapid package and a Linux Paddle package; available assets can vary by release, so check the live page.
| Choice | Start here when | Benchmark | Fallback signal |
|---|---|---|---|
| Paddle build | Modern compatible CPU and higher document volume | Pages/minute, peak RAM and accuracy | Initialization error, incompatibility or memory pressure |
| Rapid build | Older/lower-resource Windows device | Same test set and end-to-end correction time | Throughput too slow for volume |
| Linux package | Supported x64 desktop environment | Library, font, clipboard and capture behavior | Distribution-specific dependency failures |
| Docker/runtime | Controlled server or repeatable deployment | API exposure, volumes, CPU and cold start | Desktop features or hardware integration needed |
Do not choose solely from package size. Run 50 representative pages and measure character error, missed regions, processing time, memory and manual correction. The more accurate engine can be slower yet cheaper if it saves review; the faster engine can win for clean, repetitive pages.
Download and install safely
Use the official GitHub release page or mirrors named there. Compare the downloaded SHA-256 with the value published for the exact asset. The v2.1.5 release reviewed listed SHA-256 values for Windows Rapid and Linux Paddle archives. A self-extracting .7z.exe can be opened as an archive; inspect it and scan under organizational policy before execution.
Portable software still writes settings and logs. v2.1.5 added logs under UmiOCR-data/logs, with configurable saved severity. Review logs before processing secrets: file paths, errors or recognized snippets can become retained data. Protect the entire UmiOCR-data directory, backups and exported results.
Build a document-specific OCR benchmark
Create ground truth from manually verified pages. Sample the hardest conditions actually expected, not a clean screenshot. Include multiple scripts, small fonts, skew, phone photos, low contrast, stamps, handwriting, vertical text, mixed columns, tables and pages with headers/footers.
| Metric | What it catches | Why it is insufficient alone |
|---|---|---|
| Character error rate | Insertions, deletions and substitutions | One wrong digit can matter more than many punctuation errors |
| Word error rate | Word-level usability | Poor for scripts without simple word boundaries |
| Region recall | Completely missed text blocks | Does not score recognized content |
| Reading-order accuracy | Column and footnote sequencing | Needs structural ground truth |
| Critical-field accuracy | Names, totals, dates, IDs and negation | Domain-specific, but essential for decisions |
| Correction minutes/page | True workflow cost | Depends on reviewer skill and interface |
Set acceptance by use case. Search indexing may tolerate modest errors; a contract clause, medical value or invoice total may require double entry or qualified review. Never infer “99% accurate” from a different language, model or scan distribution.
Image preparation often beats model switching
Rotate to the correct orientation, crop irrelevant borders, use sufficient resolution and correct perspective on phone photos. Preserve an untouched source. Avoid aggressive thresholding that erases thin characters, decimal points or diacritics. Test grayscale, contrast and binarization on copies.
| Problem | Preprocessing experiment | Risk |
|---|---|---|
| Skewed scan | Deskew while preserving page edges | Interpolation blurs small text |
| Perspective photo | Four-corner correction | Wrong corners distort columns |
| Low contrast | Local contrast on a copy | Paper texture becomes false strokes |
| Watermark/header | Use ignore regions or post-filter rules | Rule may delete legitimate repeated text |
| Tiny characters | Rescan at higher optical resolution | Upscaling cannot recover missing detail |
| Handwriting | Use a handwriting-specialist alternative | Printed-text OCR may output confident nonsense |
Multilingual recognition
Install or select the correct language library for the document. A broad multilingual model may recognize mixed scripts but confuse visually similar characters; a narrower language model can improve accuracy when the language is known. Test Latin/Cyrillic confusables, Chinese/Japanese variants, punctuation, accents and right-to-left order.
OCR is not translation. Keep the recognized source text, then translate in a separate step with its own glossary and review. If translation uses a cloud model, the workflow is no longer fully local even though OCR was offline.
Complex layouts and the gap-tree sorter
Umi-OCR v2.1 introduced rewritten layout parsing described as a gap-tree sorting algorithm for multi-column documents. Test journals with sidebars, newspapers, forms, footnotes and captions. Visually correct recognition can still interleave columns or attach a caption to the wrong image.
- Compare plain-text order against a human reading path.
- Check whether headers and footers are excluded consistently without removing body text.
- Keep page boundaries and coordinates when downstream citation needs traceability.
- Use specialized table/document parsers when cell structure must be preserved.
Searchable double-layer PDFs
A double-layer PDF keeps page images and overlays an invisible text layer for search, copy and indexing. Umi-OCR release notes state that generation is supported from original PDFs; other imported document formats may produce text but not necessarily a rebuilt PDF. Confirm the exact version behavior.
| QA check | How | Failure consequence |
|---|---|---|
| Visual fidelity | Pixel-compare rendered pages | Archival page changes |
| Text alignment | Select/copy lines across page | Wrong nearby text copied |
| Reading order | Extract full page and compare | Screen readers/search snippets become incoherent |
| Page count/rotation | Automated and visual checks | Missing or inverted pages |
| File size | Compare original/output and compression | Unmanageable archive or degraded image |
| PDF safety | Scan attachments/scripts and use isolated viewer | Malicious input remains dangerous |
QR codes and barcodes are untrusted data
Decode first and display the literal payload; do not automatically open a URL, join Wi-Fi, send email or execute an app action. Normalize lookalike domains, block dangerous schemes and scan links independently. For inventory or payments, verify check digits and compare with a trusted record.
Formula recognition needs semantic checking
Optional plugins can convert formula images into LaTeX-like text. Check superscripts, subscripts, minus signs, multiplication symbols, matrices, brackets, Greek letters and equation alignment. Compile the output and compare visually, then have a domain expert verify meaning. A visually similar character can reverse a result.
Automation and HTTP interface security
Umi-OCR’s local HTTP capability can connect screenshot or batch recognition to another application. Bind to loopback by default, authenticate if exposed, restrict file paths and request size, and reject remote URLs unless explicitly required. Never expose a permissive OCR service directly to the internet.
Queue jobs with unique IDs, store source hashes, capture engine/plugin version and return coordinates/confidence where available. Define timeouts and quarantine corrupt images/PDFs. Use a separate account/container for untrusted documents because image and PDF decoders have their own attack surface.
Alternatives
| Option | Best fit | Tradeoff versus Umi-OCR |
|---|---|---|
| Umi-OCR | Private desktop screenshot, batch and PDF OCR | Limited collaborative/cloud document intelligence |
| PaddleOCR directly | Custom trained/deployed OCR pipeline | More engineering, wider current model access |
| Tesseract | Mature CLI/library and deterministic automation | Layout/language quality varies |
| OCRmyPDF | Command-line searchable PDF pipeline | Less desktop capture/UI functionality |
| Cloud Document AI | Forms, tables, scale and managed APIs | Upload, recurring cost and provider governance |
| Multimodal LLM | Semantic questions over a few pages | Can hallucinate and is weaker for exact extraction evidence |
| Human data entry | High-stakes fields and ambiguous scans | Higher cost but accountable verification |
Frequently asked questions
Is Umi-OCR free?
Yes. The official repository is MIT licensed. Distribution, hardware and organizational support can still create costs.
Is it fully offline?
Recognition can be local. Verify updates, plugins, synced folders, HTTP integrations and downstream translation separately.
Does it work on macOS?
The official project primarily distributes Windows and Linux builds. macOS users should compare native OCR, a VM or another supported tool.
Which build should I choose?
Benchmark Paddle first on compatible modern hardware and Rapid on lower-resource or incompatible Windows CPUs; use correction time as the final measure.
Can it make scanned PDFs searchable?
Yes, using a double-layer PDF workflow from source PDFs. Inspect hidden-text alignment and reading order before archiving.
Can OCR results be trusted automatically?
No for consequential data. Verify names, numbers, dates, negation, table cells and every decision-driving field.
Does it recognize handwriting?
Performance depends on the engine/plugin and script; use representative testing and prefer a handwriting-specialized system when required.
Primary sources
- Official Umi-OCR repository
- Official English README
- Official releases, packages and checksums
- Official Linux runtime and Docker guidance
- Official plugin catalog
- Project license
- PaddleOCR official documentation
- PaddleOCR 3.0 technical report
- Library of Congress PDF format preservation information
Last reviewed July 25, 2026. Umi-OCR engines, plugins and platform packages evolve independently. Verify the exact release asset, checksum, language model and offline behavior on your documents.