Crate monocr_onnx 

MonOcr - Mon language OCR library using ONNX models

This library provides OCR (Optical Character Recognition) functionality for Mon text using deep learning models. It supports reading text from images and PDFs, with optional accuracy measurement against ground truth text.

The model

Weights are downloaded from janakhpon/monocr, pinned to revision model_manager::MODEL_REVISION. That artifact takes a [1, 1, 160, 1024] input and emits [1, sequence, 277] logits: 276 characters plus the CTC blank. The width is static: v3.5 accepts 1024 and nothing else, where v2 accepted any width.

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 graph is read on load and a disagreement yields ModelContractError instead of a result.

Quick Start

use monocr_onnx::read_image;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let text = read_image("path/to/image.png").await?;
    println!("Recognized text: {}", text);
    Ok(())
}

Features

• Read text from single images (PNG, JPG, etc.)
• Read text from multiple images in batch
• Read text from PDF files (requires poppler-utils)
• Measure OCR accuracy against ground truth
• Customizable model paths and character sets
• Line segmentation for full page OCR
• Lines too wide for the 1024px model window are tiled at whitespace columns rather than squeezed into it; see MonOcr::predict_page for the measured reason

Re-exports

pub use model_manager::ModelManager;

pub use segmenter::cut_column;

pub use segmenter::tile_line;

pub use segmenter::CUT_INK_THRESHOLD;

pub use segmenter::CUT_SEARCH_FRACTION;

pub use segmenter::DEFAULT_DENSITY_THRESHOLD_RATIO;

Modules

model_manager

Model Manager

segmenter

Line Segmentation

Structs

BBox

Bounding box for a line or text region

LineResult

Result from line prediction

ModelContractError

A model artifact that disagrees with the charset or the input geometry this binding was built for.

MonOcr

Main OCR engine for text recognition

MonOcrBuilder

Builder for configuring and creating MonOcr instances

OcrResult

OCR result containing recognized text and accuracy measurement

Constants

DEFAULT_INPUT_WIDTH

Padded canvas width fed to the model.

EXPECTED_INPUT_HEIGHT

Input height this binding preprocesses for.

Functions

calculate_accuracy

Calculate character error rate (CER) between predicted and ground truth Returns accuracy percentage (0-100)

normalize_charset

Strip line terminators, and nothing else.

normalize_polarity

Return image as dark-text-on-light, inverting it when the background is dark.

page_text

Join one page’s line texts the way MonOcr::read_image does.

read_image

Read text from a single image file

read_image_with_accuracy

Read text from an image with accuracy measurement

read_images

Read text from multiple image files

read_pdf

Read text from a PDF file