MonOCR (Rust SDK)

A Rust SDK for Mon language OCR, powered by ONNX Runtime.

Mon (mnw) is a Mon-Khmer language of Myanmar and Thailand, written in a Myanmar-script orthography. It is unrelated to Mongolian.

Installation

Add this to your Cargo.toml:

[dependencies]
monocr-onnx = { git = "https://github.com/MonDevHub/monocr-onnx", branch = "main" }
tokio = { version = "1", features = ["full"] }

The model

Weights are downloaded from janakhpon/monocr, pinned to revision d3d9d5e (model_manager::MODEL_REVISION). That artifact takes a [batch, 1, 160, 1024] input and emits [batch, sequence, 277] logits: 276 characters plus the CTC blank. Height and width are both static; batch is the only dynamic axis. This line previously read [1, 1, 160, width], which had both halves backwards.

The charset, the input height and the classifier width are one contract. If they drift apart the model still runs and still returns text — it is just the wrong text, with no error anywhere. So the SDK reads the real graph on load and returns a ModelContractError when it disagrees with the charset it holds:

model contract violation: charset/model mismatch.
  charset: 276 characters -> expects 277 classes (276 + CTC blank)
  model (/…/monocr.onnx): 225 classes

Downloads are cached per revision under ~/.monocr/models/<revision>/, so re-pinning is a cache miss rather than a silent reuse of the previous artifact.

Features

• Auto-Model Management: Downloads and caches the pinned ONNX weights and their charset to ~/.monocr/models/<revision>/.
• Fail-closed loading: Refuses to run a model whose input height or class count disagrees with the charset, instead of returning the wrong text.
• Memory Efficient: Uses ndarray for tensor construction.
• Line segmentation: Horizontal projection profile over a flat global threshold at 128 for full-page OCR (src/segmenter.rs:288). Not adaptive — the crate's own docs at src/segmenter.rs:265 state it correctly.

Quick Start

use monocr_onnx::MonOcr;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Downloads and caches the model on first use.
    let mut ocr = MonOcr::builder().build().await?;

    let text = ocr.read_image("test_image.jpg").await?;
    println!("Recognized Text: {text}");
    Ok(())
}

Or the one-shot free functions, which build a MonOcr for you:

use monocr_onnx::{read_image, read_images, read_pdf};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let text = read_image("page.png").await?;
    let batch = read_images(&["a.png", "b.png"]).await?;
    let pages = read_pdf("document.pdf").await?; // needs poppler-utils
    println!("{} {} {}", text.len(), batch.len(), pages.len());
    Ok(())
}

Using your own model

use monocr_onnx::MonOcr;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let charset = std::fs::read_to_string("charset.txt")?;
    let mut ocr = MonOcr::builder()
        .model_path("./models/monocr.onnx")
        .charset(charset)
        .build()
        .await?;

    println!("{}", ocr.read_image("page.png").await?);
    Ok(())
}

The charset is stripped of line terminators only. Its first character is U+0020 — a space is one of the classes the model emits, so trimming it with .trim() shifts every index in the decode by one.

Prerequisites

This crate requires the ONNX Runtime shared library to be available on your system.

• macOS: brew install onnxruntime
• Linux: Download libonnxruntime.so and add to LD_LIBRARY_PATH.

License

MIT