PaddleOCR-rs

English | 中文

ONNX-based OCR engine using PaddleOCR models, written in Rust.

Features

• Text Detection & Recognition — Complete OCR pipeline with DBNet detection and CRNN recognition
• Document Orientation Classification — PP-LCNet classifier for 0°/90°/180°/270° rotation detection and auto-correction
• Hardware Acceleration — ONNX Runtime with CPU, DirectML, CUDA, OpenVINO, NNAPI, CoreML, CANN support
• Cross-platform — Windows, Linux, macOS, Android (via FFI), iOS (via FFI)
• Concurrent Processing — rayon parallel execution + session pooling
• Fine-grained Control — Step-by-step API: detect → recognize, or full pipeline with ordering modes
• Minimal Dependencies — Only ONNX models needed, no external runtime libraries

Quick Start

use paddleocr_rs_onnx::{OcrEngine, OrderBy};

let det_model = std::fs::read("ch_PP-OCRv4_det_infer.onnx")?;
let rec_model = std::fs::read("ch_PP-OCRv4_rec_infer.onnx")?;
let keys = std::fs::read("ppocr_keys_v1.txt")?;

let engine = OcrEngine::new(&det_model, &rec_model, &keys)?;

let image = image::open("test.png")?;
let blocks = engine.recognize_all(&image, OrderBy::Horizontal)?;
for block in &blocks {
    println!("{} ({:.2}%)", block.text, block.confidence * 100.0);
}

Installation

[dependencies]
paddleocr_rs_onnx = "0.2"

API Documentation

• API Overview — Modules, structs, return types
• OcrEngine — Detection and recognition API
• DocOrientationClassifier — Orientation detection and correction
• Hardware Acceleration — AccelerationDevice, platform requirements, EP features
• FFI API — C-compatible API

Mobile Platform Support

PaddleOCR-rs supports Android and iOS platforms via C FFI interface.

Android

• CPU: ✅ Supported (via ONNX Runtime)
• NNAPI: ✅ Supported (via nnapi feature)
• Targets: aarch64-linux-android, armv7-linux-androideabi (requires custom build), x86_64-linux-android (requires custom build)

iOS

• CPU: ✅ Supported (via ONNX Runtime)
• CoreML: ✅ Supported (via coreml feature)
• Targets: aarch64-apple-ios, aarch64-apple-ios-sim

Building for Mobile

# Android
./build-android.sh aarch64-linux-android --release

# iOS
./build-ios.sh aarch64-apple-ios --release

Example Projects

• xc-ocr-onnx — Tauri 2 desktop GUI app, supports image / DOCX / PDF OCR with dynamic model switching
• examples/android-demo/ — Android Kotlin example
• examples/ios-demo/ — iOS Swift example

Comparison with Other Rust PaddleOCR Implementations

This project is one of several Rust implementations of PaddleOCR. Below is a comprehensive comparison of the three main implementations:

Acceleration Hardware Support Comparison

Platform/Backend	PaddleOCR-rs	paddle-ocr-rs	rust-paddle-ocr
Windows
CUDA	✅ (via cuda feature)	✅	✅
DirectML	✅	✅	❌
OpenVINO	✅ (via openvino feature)	❌	❌
Linux
CUDA	✅ (via cuda feature)	✅	✅
CANN	✅ (via cann feature)	✅	❌
OpenVINO	✅ (via openvino feature)	❌	❌
macOS
Metal	✅ (via metal feature)	❌	✅
CoreML	✅ (via coreml feature)	❌	✅
Android
NNAPI	✅ (via nnapi feature)	❌	❌
CPU	✅	✅	✅
iOS
CoreML	✅ (via coreml feature)	❌	✅
CPU	✅	✅	✅

Comprehensive Comparison

Feature	PaddleOCR-rs	paddle-ocr-rs	rust-paddle-ocr
Model Format Support	✅ ONNX only	✅ ONNX format	✅ MNN format
Document Orientation	PP-LCNet classifier	PP-OCR v2.0 classifier	PP-LCNet classifier
Concurrency	✅ rayon parallel + session pooling	✅ rayon parallel + batch inference	⚠️ rayon in pre/post-processing, inference is single-threaded
External Interfaces	✅ Rust API + C FFI API (via ffi feature)	✅ YAML config + CLI (rapidocr)	✅ C API (cdylib) + CLI (newbee-ocr-cli)
Memory/Type Safety	✅ Memory-safe Rust + strong typing	✅ Memory-safe Rust + strong typing	✅ Memory-safe Rust (mnn-rs) + ⚠️ C API partial
Concurrency Safety	✅ Thread-safe by design	✅ Thread-safe (Arc + Mutex)	⚠️ Requires careful handling

Thanks

This project is built upon the work of the following projects:

• PaddleOCR - Provides models
• FastDeploy - Provides runtime reference
• MaaFramework - Provides architecture reference

License

MIT