Files
ssw c303c315e5 rocr: rusto-rs 源码入库(上游 v0.2.5 基线 57609fc + 本地补丁)
- 源码 316 文件收编为 rusto-rs/ 文件夹,随本仓库管理
- 移除内部 .git(补丁已存档 docs/rusto-rs-metal-patch-20260912.patch,上游可随时 re-clone)
- target/ 与 docs/.yarn 缓存不入库;本地改动:ROCR_* 环境变量调优 + Metal 链接修复 + 预处理测试脚本
2026-09-12 17:30:04 +08:00

6.0 KiB

title, description, navigation, seo
title description navigation seo
Introduction High-Performance Pure Rust OCR Library & Multi-Platform Toolkit
icon
i-ph-rocket
title description
Introduction - RustO! High-Performance Pure Rust OCR Library & Multi-Platform Toolkit

::project-badges ::

RustO! is a high-performance Optical Character Recognition (OCR) engine and cross-platform toolkit written in pure Rust. Based on RapidOCR and powered by PaddleOCR deep learning models with Alibaba's MNN lightweight inference backend, RustO! delivers sub-second inference speeds, ultra-low memory overhead, zero OpenCV runtime dependencies, and 99.3%+ accuracy parity with native C++ OpenCV implementations.

🎯 Architecture Pipeline

RustO! orchestrates an end-to-end OCR and layout analysis pipeline optimized for edge devices and high-throughput servers:

flowchart TD
    A["Input Image Source<br/>(Filesystem Path, Bytes Buffer, or Native URI)"] --> B["1. Preprocessing & Bounded Resize<br/>(Aspect-Ratio Scaling, Padding, Normalization)"]
    B --> C["2. DBNet Text Detection<br/>(Heatmap, Binarization, Dilation, Contour Unclipping)"]
    C --> D{"3. Line Orientation CLS?<br/>(Optional: Inverted 0° vs 180°)"}
    D -->|Enabled| E["Rotate Upside-Down Text 180°"]
    D -->|Bypassed / Disabled| F["Perspective Crop & Rayon Parallel Batching"]
    E --> F
    F --> G["4. CTC Text Recognition<br/>(Batched SVTR/CRNN, 50-Language Dictionary)"]
    G --> H{"Output Granularity?"}
    H -->|lines / words| I["Structured Output<br/>(TextResult with BoxPoints & Frame)"]
    H -->|spatial| J["5. Spatial Layout Reconstruction<br/>(Row Y-Clustering & Column X-Alignment)"]
    J --> K["Formatted Spatial Text<br/>(Preserved Tables, Receipts, Indentations)"]

⚡ Key Highlights & Advantages

  • 🚀 100% Pure Rust Core: Zero OpenCV runtime dependency. Binarization, contour tracing, polygon clipping, and affine transformations are implemented in memory-safe Rust with SIMD acceleration.
  • 📦 Ultra-Small Footprint: Core binaries are ~5 MB (self-contained) compared to 50+ MB for traditional OpenCV + ONNX runtime stacks.
  • 📄 Human-Readable Spatial Layouts: Native support for reconstructing invoices, receipts, multi-column tables, and forms directly into visually formatted text.
  • 🧠 Comprehensive Model Ecosystem: Out-of-the-box support for PP-OCRv6 (Tiny, Small, Medium), PP-OCRv5 (Mobile, Server), and PP-OCRv4 (Mobile, Server) architectures.
  • 🌐 50+ Languages in One Model: PP-OCRv6 uses a unified dictionary supporting Latin, Cyrillic, CJK, Devanagari, Arabic, and more simultaneously.
  • 📱 Universal Multi-Platform SDKs: First-class support for Rust, .NET / C#, React Native, iOS (Swift), Android (Kotlin), and C FFI.

📦 Supported Platforms & Ecosystem

Platform Target Package / Registry Support Matrix
Rust cargo add rusto-rs (crates.io) Linux, macOS, Windows, Android (NDK), iOS
.NET / C# dotnet add package RustODotnet (NuGet) .NET 6.0+, .NET 8.0+, .NET Standard 2.0+ (Win/Mac/Linux)
React Native npm install react-native-rusto (npm) React Native 0.70+ (New Arch & Paper), Expo
iOS pod 'RustO' (CocoaPods) iOS 13.0+, Mac Catalyst (arm64, x86_64)
Android com.github.byrizki.rusto-rs:rusto-android (JitPack) Android API 21+ (ARM64-v8a, ARMv7a, x86_64, x86)
C / Native librusto.so / librusto.dylib / rusto.dll C99 / C++11 and any language with C FFI bindings
CLI rusto binary Standalone command-line tool for desktop & CI pipelines

📊 Performance & Parity Comparison

Benchmarks conducted on standard document scans (A4 invoice, 150 DPI) on an AMD Ryzen 7 / Apple M-series CPU:

Metric RustO! (Pure Rust + MNN) Traditional OpenCV + C++ Pipeline Advantage
Detection Speed ~80 ms ~85 ms ⚡ 6% faster
Recognition Speed ~120 ms ~125 ms ⚡ 4% faster
Accuracy Parity 99.3%+ Baseline (100%) 🎯 Identical bounding boxes
Binary Footprint ~5 MB ~55 MB+ (libopencv, onnxruntime) 📦 90% reduction
Peak Memory (RAM) ~120 MB ~260 MB+ 🔒 54% lower memory
Safety Memory-safe (Rust guarantees) Manual pointers / potential leaks 🛡️ Zero memory bugs

Next Steps