Multilingual OCR

RustO! provides OCR capabilities across dozens of languages. To achieve high accuracy while keeping model sizes small, models are divided into core bilingual models (PP-OCRv6) and specialized regional language packs (PP-OCRv5 & PP-OCRv4).

1. PP-OCRv6 Language Coverage (Chinese + English / Latin)

The default PP-OCRv6 models are trained primarily for Chinese and English, with broad coverage for basic Latin-alphabet languages:

Important: PP-OCRv6 does not recognize Korean Hangul, Arabic, Cyrillic, Devanagari (Hindi), Thai, Greek, or South Indian scripts. For those writing systems, use the dedicated language models below.

2. Dedicated Regional Language Packs (PP-OCRv5 & PP-OCRv4)

Because text detection (det.rten) is language-agnostic (it detects the physical contours of text lines regardless of language), you can recognize any supported writing system by pairing the universal detector with a language-specific recognizer (rec.rten) and dictionary (dict.txt).

Base release URL: https://github.com/byrizki/rusto-rs/releases/download/models-rten-v0.3.2/

3. How to Switch Language Models in Code

To switch languages, keep the universal text detector (det.rten) and point the recognition configuration to your downloaded regional model:

In Rust:

use rusto::{InitializeConfig, RustO};

// Initialize with universal detector + Japanese recognition model
let config = InitializeConfig::ppv4(
    "models/det.rten",
    "models/japan_rec.rten",
    "models/japan_dict.txt",
);

let mut ocr = RustO::initialize(config)?;
let results = ocr.detect_text_file("japanese_receipt.jpg")?;

In Node.js:

import { RustO } from '@rustors/node';

const ocr = new RustO({
  detModelPath: 'models/det.rten',
  recModelPath: 'models/arabic_rec.rten',
  dictPath: 'models/arabic_dict.txt',
});

const lines = await ocr.detectTextFile('arabic_invoice.pdf');

In Python:

import rusto

engine = rusto.RustOEngine.from_config(
    '{"detection":{"model_path":"models/det.rten"},"recognition":{"model_path":"models/korean_rec.rten","dict_path":"models/korean_dict.txt"}}'
)

lines = engine.detect_text_file("korean_document.png")

4. Package Manager Identifiers

If using CocoaPods, Android Gradle, or NuGet, pre-packaged language pods/AARs/packages are available without manual downloads:

5. Right-to-Left (RTL) Scripts

When processing Arabic, Persian, or Urdu:

• The recognizer reads characters in their natural visual order from left to right across the cropped text slice.
• Bounding polygon vertices in structured output reflect true coordinate space on the document.
• In spatial text mode (OutputGranularity::Spatial), lines are formatted preserving individual word coordinates and token ordering.