Images Translator

English · عربي

Turn scanned pages — manuals, brochures, spec sheets — into translated pages that still look like the original. The tool reads the text with OCR, translates it, and redraws the translation right on top of the scan, keeping the tables, diagrams, icons and logos exactly where they were. Then you fine-tune the result in a visual editor.

Think of it as Google Translate's camera mode, but full-resolution, editable, scriptable, and able to run fully offline. Works on macOS, Linux and Windows.

A Japanese manual, automatically translated to English — layout, tables and artwork untouched.

What it can do

• Read 30+ languages with the macOS Vision OCR engine, or Tesseract (works on older macOS, Linux and Windows too — pick it in Settings).
• Translate offline with Apple's built-in translator (macOS 15+), or through Google Cloud Translation — including right-to-left targets like Arabic, with correct shaping and alignment.
• Edit everything visually in your browser: move, resize and rewrite text boxes, change fonts, colors and backgrounds, pick a patch color straight from the image, undo with ⌘Z. What you see is exactly what gets saved.
• Save the way you want: PNG, optimized PNG-8, grayscale, crisp 1-bit black & white, JPEG — or a per-page PDF with a hidden, searchable text layer.
• Batch a whole book from the command line; every page is cached, so re-runs take seconds.

Samples

Pages of a 1990s Japanese car-manual, straight out of the tool.

Example 1:

• Original scanned page (Japanese)
• Translated to English — PDF with a searchable text layer
• Translated to Arabic — optimized PNG-8

Example 2:

• Original scanned page (Japanese)
• Translated to English — optimized PNG-8

Installation

Runs on macOS, Linux and Windows. You need Python 3 and, for OCR, either macOS Vision (macOS 15+) or Tesseract (everywhere else).

git clone https://github.com/techana/scanned-images-translator.git
cd scanned-images-translator
pip3 install -r requirements.txt

Then set up OCR for your platform:

macOS 15+ — the built-in Vision engine is the default; just build the two small Swift helpers once (needs Xcode CLT — xcode-select --install):

swiftc -O -o ocr ocr.swift          # Vision OCR
swiftc -O -o translate translate.swift   # Apple offline translation

Offline translation then downloads its language pack on first use.

Linux — install Tesseract, its language data, and fonts:

sudo apt install tesseract-ocr tesseract-ocr-jpn tesseract-ocr-ara \
                 fonts-dejavu fonts-noto zenity
# (swap in the tesseract-ocr-<lang> packages for your source languages)

Windows — install Tesseract (the UB-Mannheim build; tick the language data you need during setup) and make sure tesseract.exe is on your PATH. Fonts and file dialogs are already present.

On Linux/Windows the app defaults to Tesseract for OCR and Google for translation (Apple's offline translator is macOS-only). Pick the OCR engine, languages and translator any time in Files ▾ → Settings…

Using it

On a Mac, just double-click Images Translator.app in this folder — no Terminal needed. It starts the app and opens it in your browser. To stop it, choose Files ▾ → Quit.

First time only: if macOS says the app "cannot be opened because it is from an unidentified developer", right-click it once and choose Open → Open. (That only happens if you downloaded the project as a ZIP; a git clone isn't flagged.)

Or start it from a terminal — on any platform:

python3 gui.py

The app runs as a small local web server and opens automatically in your browser at http://localhost:8877 — if it doesn't open, just paste that address into any browser. (Use python3 gui.py --port 9000 if something else is already using port 8877.) Nothing is uploaded anywhere; the server only listens on your own machine.

Then Files ▾ → Load a folder…, pick your scans, and pages translate as you view them. Click any text box to adjust it; drag the Opacity slider to peek at the original underneath. ⌘S saves the page — the first save asks where and in which format, and remembers your answer.

Languages and the translation engine live in Files ▾ → Settings…

The command line, for whole folders at once:

python3 translate_pages.py scans/                          # Japanese → English
python3 translate_pages.py --source ja-JP --target ar scans/   # → Arabic

Output lands next to the inputs as <page>_<language>.png.

Good to know

• A recent Mac gives the best results. macOS Vision (the default on macOS 15+) runs Apple's neural text-recognition models, trained on real-world photos and scans, and it clearly outperforms Tesseract on noisy halftone pages — especially Japanese. Tesseract is neural too (LSTM, since v4) and is a solid portable fallback for Linux, Windows and older macOS, but expect rougher OCR on grainy scans, which means more fixing up by hand in the editor.
• Copy and paste elements with ⌘C / ⌘V (Ctrl on Linux and Windows). Works on the selected text box or patch rectangle: the duplicate keeps everything — text, size, font, colours, alignment — and lands wherever the mouse is pointing. Handy for repeating a caption style down a page.
• OCR and translation results are cached per page in images_translator.cache/ — edits live there too, and re-rendering a full book takes seconds.
• ./ocr --list-langs and ./translate --list-langs show every language your machine supports.
• Recurring OCR mistakes in a specific document can be fixed once with a small corrections file (see fixes-carmarty.json for an example, and pass it with --fixes).

License

MIT