OCR — Pure Rust Optical Character Recognition

A from-scratch OCR system built in Rust, informed by decades of OCR research and modern deep-learning pipelines. The goal is a clean, extensible, pure-Rust toolchain that can rival Tesseract on CPU and approach EasyOCR/PaddleOCR accuracy.

Current State

Compiles and passes 357+ tests. The full pipeline works end-to-end for printed text on clean images, with a modular architecture supporting both classical and learned recognition:

• Image preprocessing: deskew, binarization (Otsu, Sauvola), noise reduction, contrast enhancement, auto-rotate (0°/90°/180°/270°)
• Layout analysis: Union-Find CCL, column/line detection, reading-order resolution, table detection, form field extraction
• Text detection: TextDetector trait with CCL and lightweight CNN implementations
• Recognition: Pattern-matching engine (trainable from synthetic fonts) + CRNN (CNN + BiLSTM + CTC) selectable via --engine
• Multi-language: Unicode script detection (Latin, CJK, Arabic, Cyrillic, Greek, Hebrew, Thai, Devanagari), dictionaries for 25+ languages, per-script CRNN vocabularies
• Post-processing: Dictionary-based spell correction, document structure classification (headings, lists, paragraphs), hierarchical Markdown/JSON output
• Output formats: text, JSON, hOCR, TSV, ALTO XML, box files, searchable PDF (invisible text overlay), Markdown, structured JSON
• CLI: extract, batch, layout, list-languages, check, info, validate
• Web API (feature web-api): HTTP server with multipart upload
• PDF input (feature pdf)

What We Learned from Popular OCR Offerings

Key Architectural Lessons

1. Two-stage pipeline (detection → recognition) dominates for documents. End-to-end VLMs are promising but overkill for pure OCR.
2. CTC decoders remove the need for per-character segmentation in sequence models.
3. Synthetic training data (rendered fonts + realistic distortions) lets you bootstrap accuracy before collecting real data.
4. Image preprocessing matters: deskew, binarization, and contrast enhancement can improve accuracy 10–20%.
5. Post-processing dictionaries / language models close the gap between raw recognition and human-level accuracy.
6. CPU deployment still matters: Tesseract remains ubiquitous because it runs everywhere without GPU dependencies.

Roadmap

Phase 0 — Baseline (DONE)

• End-to-end pipeline: image → preprocess → layout → recognize → output
• Pattern-matching recognition engine for Latin glyphs
• Layout analysis with CCL, column detection, reading order
• Dictionary-based post-correction
• CLI with extract, batch, layout, list-languages
• 357+ tests passing

Phase 1 — Synthetic Training Infrastructure (DONE)

• Synthetic text-image generator with TTF font rendering and bitmap fallback
• Distortion pipeline (rotation, blur, noise, contrast, shear)
• CER/WER benchmark harness with clean/mild/heavy test sets
• Train pattern-matching templates from synthetic renders (TemplateTrainer)

Phase 2 — Robust Text Detection (DONE)

• TextDetector trait with CclDetector and CnnDetector
• Lightweight 3-layer detection CNN in pure Rust (ndarray)
• Synthetic document generator + IoU-based evaluation harness
• Auto-rotate 0°/90°/180°/270° via projection-variance orientation detection

Phase 3 — Learned Recognition (CRNN) (DONE)

• CNN feature extractor (5 conv layers + maxpool)
• 2-layer bidirectional LSTM (64 hidden)
• CTC decoder (greedy + beam search)
• CTC loss with forward-backward algorithm
• CrnnTrainer with synthetic data + checkpoint saving
• Wired into OcrEngine via --engine lstm
• Model size < 5MB (~2.4MB default config)

Phase 4 — Multi-Language Scale (DONE)

• Unicode script detection and routing (8 scripts)
• Dictionaries for 25+ languages
• Multi-script synthetic data generation (ScriptLineGenerator)
• Per-script CRNN models (ScriptModelRegistry)

Phase 5 — Advanced Layout & Structure (DONE)

• Table detection with grid line scan and cell reconstruction
• Row/column span inference via pixel-density checks
• Form field extraction (checkboxes, key-value pairs, underline fields)
• Document structure classification (headings, paragraphs, lists, captions)
• Hierarchical Markdown and structured JSON output
• Searchable PDF generation with invisible text overlay

Next Steps (requires training execution)

• Train CRNN to CER < 5% on clean synthetic test
• Train CRNN to CER < 15% on distorted synthetic test
• Evaluate per-language accuracy on synthetic benchmarks

Installation

# Default build (pattern matching engine)
cargo build --release

# With all features
cargo build --release --all-features

# Run tests
cargo test

Usage

# Recognize text from an image
ocr extract document.png

# JSON output with bounding boxes and confidence
ocr extract document.png -f json

# Markdown output with document structure
ocr extract document.png -f markdown

# Structured JSON with headings, paragraphs, lists
ocr extract document.png -f structured-json

# Enable preprocessing pipeline
ocr extract document.png --preprocess

# Use CRNN engine
ocr extract document.png --engine lstm

# Batch process a directory
ocr batch -i ./images -o ./results --lang en

# List supported languages
ocr list-languages

# Analyze image layout
ocr layout document.png -o layout.json

Library

use ocr::api::Ocr;
use ocr::core::config::OcrConfig;

let config = OcrConfig::default();
let ocr = Ocr::with_config(config)?;
ocr.initialize().await?;

let result = ocr.recognize_text_from_file("document.png").await?;
println!("{}", result.text);

Project Structure

src/
├── api/          High-level OCR API (Ocr, TextProcessor)
├── cli/          CLI argument parsing with clap
├── core/         Core OCR engine (config, engine, geometry, layout, output, text)
├── image/        Image preprocessing pipeline (binarization, enhancement, quality)
├── lang/         Language support (CJK, detection, dictionary, N-gram, unicode)
├── layout/       Layout analysis (column/line detection, text ordering, CCL, detection CNN, table/form extractors, classifier)
├── recognition/  Recognition models (pattern matching, CRNN, CTC decoder, LSTM)
├── synthetic/    Synthetic data generation (fonts, distortion, multi-script, document layouts, template training)
├── training/     Model training pipeline (data, losses, metrics, optimizers, CRNN trainer)
└── utils/        Shared utilities (async, error, hash, math, SIMD, time)

License

Apache-2.0