trivision-ocr 1.0.0

Multi-engine OCR pipeline — beats Google Vision API

ocr-pipeline

Multi-engine OCR pipeline combining Tesseract, PaddleOCR, and EasyOCR with confidence-weighted voting and domain-aware spell correction. Output is compatible with the Google Vision API JSON schema.

Architecture

Image → [Preprocessor] → Tesseract ┐
                       → PaddleOCR ├─→ [Voter] → [Corrector] → Vision API JSON
                       → EasyOCR   ┘

Stage	What it does	Your edge vs Vision API
0 — Preprocessor	Hough deskew, CLAHE, binarize, optional 2× EDSR	Domain-specific prep; Vision API gets raw images
1 — Three engines	Tesseract (LSTM), PaddleOCR (DBNet+SVTR), EasyOCR (CRAFT+CRNN)	Different failure modes → ensemble eliminates each
2 — Voter	IoU spatial grouping + weighted confidence + agreement bonus	Cross-engine consensus exposed; Vision API hides this
3 — Corrector	SymSpell + domain vocabulary protection	Tuned to your label vocabulary; Vision API uses general LM

Prerequisites

Tesseract binary (Windows — required):

winget install UB-Mannheim.TesseractOCR

Python 3.10+ must be installed and on PATH.

Quick Start

# 1. Clone / copy the project
cd ocr-pipeline

# 2. Create virtual environment and install (CPU default)
setup_venv.bat

# For GPU (CUDA + paddlepaddle-gpu):
setup_venv.bat gpu

# 3. Activate
.venv\Scripts\activate

# 4. Run on an image
ocr-pipeline path\to\image.jpg --pretty

# Save output to file
ocr-pipeline path\to\image.jpg --output result.json

# Text only
ocr-pipeline path\to\image.jpg --text-only

# Skip super-resolution (faster, no EDSR model needed)
ocr-pipeline path\to\image.jpg --no-super-resolve

Python API

from ocr_pipeline import beat_vision_api

result = beat_vision_api("path/to/image.jpg")
print(result["responses"][0]["fullTextAnnotation"]["text"])

Configuration

Copy .env.example to .env and edit:

TESSERACT_CMD=C:\Program Files\Tesseract-OCR\tesseract.exe
USE_GPU=false
USE_SUPER_RESOLVE=false

Add project-specific label codes to DOMAIN_TERMS in ocr_pipeline/config.py.

Running Tests

pytest tests/test_voter.py -v

All 7 tests run without GPU or model downloads.

Optional: Super-Resolution

EDSR 2× upscale (~143 MB model) significantly improves small text on diagram scans. Enable it:

1. Run python download_models.py and choose y when prompted for EDSR
2. Set USE_SUPER_RESOLVE=true in .env

Project Structure

ocr-pipeline/
├── ocr_pipeline/       ← Python package
│   ├── config.py       ← All tunable parameters
│   ├── preprocessor.py ← Stage 0
│   ├── engines.py      ← Stage 1
│   ├── voter.py        ← Stage 2
│   ├── corrector.py    ← Stage 3
│   └── pipeline.py     ← Orchestration
├── tests/
│   └── test_voter.py   ← Unit tests (no GPU)
├── models/             ← Downloaded model files
├── setup_venv.bat      ← Windows setup
└── download_models.py  ← Model downloader