OCR Translator

Extract text from images with Tesseract, translate it, and keep a searchable history of every run — from a single file, a URL, or a live camera feed.

---

Why preprocessing matters more than flags

Tesseract's accuracy is dominated by what you hand it. A slightly rotated phone photo of a page, or one with low contrast, can drop from readable to unusable — and no --psm setting recovers it. This project puts the preprocessing pipeline first and makes it measurable: compare runs the same image under every profile and prints confidence side by side, so you pick a profile from evidence rather than guesswork.

$ python cli.py compare scan.jpg --lang eng
profile        chars   confidence   seconds
light            ...          ...       ...
standard         ...          ...       ...
aggressive       ...          ...       ...
document         ...          ...       ...

Pick the profile with the highest confidence and a plausible character count — a profile that scores well but returns far fewer characters has thrown away text.

Preprocessing profiles

Profile	Contrast	Denoise	Deskew	Binarize	Resize	Use for
light	yes	yes	no	no	1.0x	clean screenshots
standard	yes	yes	yes	no	1.2x	general photos
aggressive	yes	yes	yes	yes	1.5x	small or faint text
document	yes	yes	yes	yes	1.0x	scanned pages

Deskew

Skew correction uses the projection-profile method: the image is binarised, rotated through a range of candidate angles, and each rotation is collapsed into a horizontal projection of row sums. When text lines sit horizontally the projection swings sharply between dense text rows and empty gaps, so its variance peaks — that angle is the skew.

Verified against synthetic pages with known rotation:

True skew	Estimated
0.0°	0.0°
+5.0°	+5.0°
-8.0°	-8.0°
+11.0°	+11.0°

Quickstart

1. Install Tesseract

The Python package is a wrapper — the Tesseract binary must be installed separately.

• Windows: install from UB Mannheim's build, then either add it to PATH or pass --tesseract "C:\Program Files\Tesseract-OCR\tesseract.exe"
• macOS: brew install tesseract
• Linux: sudo apt install tesseract-ocr

Language packs are separate. For Persian: sudo apt install tesseract-ocr-fas (or tick the language during the Windows install).

2. Install and run

git clone https://github.com/MrDanial-Rafiee/ocr-translator.git
cd ocr-translator
pip install -r requirements.txt

python cli.py image examples/sample_image.jpg --lang eng --to fa

Usage

# One image or a URL
python cli.py image page.jpg --lang eng --to fa --profile document
python cli.py image https://example.com/sign.png --lang eng+fas

# Skip preprocessing entirely
python cli.py image clean_scan.png --profile none

# Which profile suits this image?
python cli.py compare scan.jpg --lang eng

# Live camera
python cli.py camera --ip 192.168.1.3 --lang eng      # phone running an IP-webcam app
python cli.py camera --device 0 --lang eng            # local webcam

# History and export
python cli.py history --limit 20
python cli.py stats
python cli.py export results.csv

As a library

from ocr_translator import OCRTranslator

ocr = OCRTranslator()
result = ocr.ocr_and_translate("page.jpg", dest_lang="fa", languages=["eng"])

print(result["original_text"])
print(result["translated_text"])
print(result["ocr_confidence"])

ocr_and_translate accepts a path, a URL, a PIL.Image, or an OpenCV BGR array — the camera module passes frames straight through.

Camera controls

Key	Action	Key	Action
t	OCR + translate the frame	f	face detection
c	save the frame	m	motion detection
s	overlay the last OCR text	g	grayscale
+ / -	zoom	e	edge detection
h	recent history	l	mirror
d	statistics	q	quit

The camera address defaults to 192.168.1.3:8080 and can be set with --ip / --port or the CAMERA_IP / CAMERA_PORT environment variables.

Storage

Every run is recorded, so results are reviewable long after the fact:

ocr_translator_results/
├── results.db              SQLite: run history + translation cache
├── json_results/           full result per run, with metadata
├── text_results/           plain text: extracted + translated
└── processed_images/       what Tesseract actually saw

The translation cache is keyed on (text, source language, target language), so re-running the same image never pays for the same translation twice.

Project structure

├── cli.py                      command-line entry point
└── ocr_translator/
    ├── preprocessor.py         profiles, contrast/denoise/binarize, projection-profile deskew
    ├── translator.py           Tesseract OCR + translation + confidence
    ├── storage.py              SQLite history, translation cache, CSV export
    └── camera.py               live IP-webcam / webcam front end

Limitations and next steps

• googletrans is an unofficial client for a private Google endpoint. It breaks whenever that endpoint changes and is rate-limited without warning. A paid translation API, or a local model such as argos-translate, would make this dependable. This is the weakest link in the project.
• Confidence is Tesseract's own word confidence, averaged. It correlates loosely with correctness and should not be read as an accuracy figure.
• No accuracy benchmark. Scoring the profiles against a ground-truth set (character error rate on a labelled corpus) would turn the compare output from relative to absolute.
• Deskew searches ±15° in 0.5° steps. Heavily rotated or perspective-distorted photos need a four-point transform, which is not implemented.
• No layout analysis. Multi-column pages and tables come out as running text; --psm 6 assumes a single uniform block.
• The camera module needs a display and cannot run headless.

License

MIT — see LICENSE.