0.1.0 — 2026-07-26

Release Notes

First release. deepocr scan.pdf writes scan.ocr.pdf: the original page image with an
invisible, selectable text layer on top — searchable in any PDF viewer, produced entirely on
your machine.

Added

• Searchable PDFs. Recognized words are written in text render mode 3 over the page image,
  positioned on the words themselves, so selection, Ctrl-F and copy all land where they should.
  The font is a generated glyphless TrueType (~600 bytes) addressed through Identity-H, with an
  identity ToUnicode map so copied text comes back as the recognized text.
• Inputs: scanned PDFs, born-digital and mixed PDFs, PNG, JPEG, and multi-page TIFF (one page
  per frame). A scanned page is lifted out of the PDF without re-rendering; anything else is
  rasterized with hayro at --dpi.
• Scans are preserved byte for byte. A page that arrived as an encoded stream goes into the
  output unchanged, so a JPEG scan loses nothing to a round trip. Only pages that had to be
  rasterized are encoded, once.
• --deskew measures a page's tilt with a projection profile and rotates it back.
  --rotate auto recognizes a thumbnail four ways up and keeps the orientation that reads;
  the output page is built upright without resampling the scan.
• --optimize recompresses page images (with --quality), downsampling to --dpi where a
  scan is finer than that. It never makes a page larger — if recompression does not win, the
  original stream is kept.
• --pages limits which pages are recognized; every page of the input still reaches the
  output. --skip-text (default), --force and --redo decide what happens to pages
  that already carry text.
• --sidecar writes the plain text beside the PDF, pages separated by a form feed.
• Batch: --recursive walks a folder and mirrors its tree into --output. One unreadable
  file is reported and counted, not fatal; the run ends with a summary.
• Offline by default. Release binaries (Homebrew, GitHub Releases, the installers) carry the
  OCR models inside them — nothing is downloaded, and no document leaves the machine. A
  cargo install build fetches the models into a cache directory on first run, verifying them
  against pinned checksums; --model / DEEPOCR_MODEL_DIR override the location for air-gapped
  setups.
• deepocr-core is published alongside the binary, so the engine can be embedded elsewhere.

Known limitations

• Latin script only, which is what ocrs reads today. Other scripts arrive with additional
  backends behind the same OcrBackend trait.
• A page skipped by --skip-text keeps its image but not its text layer. The output is
  rebuilt from page rasters, so an existing layer is not carried across; --redo recognizes such
  pages again. Adding the layer to the original page in place is planned.
• Output is a plain searchable PDF, not PDF/A.

Notes

• No Ghostscript, no Python, no runtime dependencies — one binary. Every crate in the dependency
  graph is permissively licensed (MIT / Apache-2.0 / BSD / MPL / Zlib).
• The OCR models come from ocrs-models and are
  mirrored in this repository's models-v1 release so that a pinned checksum stays valid.

Install deepocr 0.1.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/deeplabua/deepocr/releases/download/v0.1.0/deepocr-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/deeplabua/deepocr/releases/download/v0.1.0/deepocr-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install deeplabua/tap/deepocr

Download deepocr 0.1.0

File	Platform	Checksum
deepocr-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
deepocr-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
deepocr-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
deepocr-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

OCR models v1 (ocrs mirror)

Mirror of the Latin OCR models DeepOCR pins, so fetching them does not depend on an unversioned upstream URL.

The files are byte-for-byte the ones published at https://ocrs-models.s3-accelerate.amazonaws.com/, and match the checksums pinned in crates/core/src/models.rs:

file	sha256
text-detection.rten	f15cfb56bd02c4bf478a20343986504a1f01e1665c2b3a0ad66340f054b1b5ca
text-recognition.rten	e484866d4cce403175bd8d00b128feb08ab42e208de30e42cd9889d8f1735a6e

This tag never moves. A new set of weights gets a new tag, so a pinned checksum stays valid forever.

Where the models come from

They are produced by robertknight/ocrs-models and are the ones the ocrs engine downloads on first run. Per that project's README they are trained exclusively on open datasets with non-restrictive licences (including HierText, CC-BY-SA 4.0).

The upstream repository does not state a licence for the model files themselves; we have asked and will record the answer here. The ocrs engine code is MIT OR Apache-2.0, as is DeepOCR.