oar-ocr-core

Foundational types, model abstractions, and task-specific predictors for the OAR OCR library.

oar-ocr-core is the engine room of the OAR OCR ecosystem. It provides the core traits and implementations for high-performance OCR pipelines, featuring ONNX-based inference, specialized image processing, and a decoupled architecture designed for extensibility and speed.

Architecture

This crate implements a three-layer architecture to ensure modularity and maintainability:

1. Models: Low-level wrappers around ONNX Runtime sessions, handling raw tensor input/output.
2. Adapters: Traits and implementations that bridge raw model outputs to domain-specific types, handling pre- and post-processing logic.
3. Tasks: Semantic contracts that define what a predictor does (e.g., "Text Detection"), ensuring consistent APIs across different model implementations.

Installation

Add oar-ocr-core to your project:

cargo add oar-ocr-core

Feature Flags

The root oar-ocr crate forwards the same feature names. See the Cargo feature guide for provider selection, system requirements, default behavior, and recommended combinations.

Quick Start

Text Detection

Detect text regions in an image using a DBNet-based model:

use oar_ocr_core::predictors::TextDetectionPredictor;
use oar_ocr_core::utils::load_image;

// 1. Initialize the predictor
let predictor = TextDetectionPredictor::builder()
    .build("pp-ocrv5_mobile_det.onnx")?;

// 2. Load and process
let image = load_image("document.jpg")?;
let results = predictor.predict(vec![image])?;

// 3. Access results (detections for the first image)
for det in &results.detections[0] {
    println!("Box: {:?}, Score: {:.2}", det.bbox, det.score);
}

Text Recognition

Recognize text from cropped image regions:

use oar_ocr_core::predictors::TextRecognitionPredictor;
use oar_ocr_core::utils::load_image;

let predictor = TextRecognitionPredictor::builder()
    .dict_path("ppocrv5_dict.txt")
    .build("pp-ocrv5_mobile_rec.onnx")?;

let image = load_image("text_line.jpg")?;
let results = predictor.predict(vec![image])?;

// Recognition returns results per input image
for (text, score) in results.texts.iter().zip(&results.scores) {
    println!("Text: {}, Confidence: {:.2}", text, score);
}

Layout Analysis

Analyze the structure of a document to identify titles, tables, and figures:

use oar_ocr_core::predictors::LayoutDetectionPredictor;
use oar_ocr_core::domain::LayoutDetectionConfig;
use oar_ocr_core::utils::load_image;

let predictor = LayoutDetectionPredictor::builder()
    .model_name("pp-doclayoutv2")
    .with_config(LayoutDetectionConfig::with_pp_doclayoutv2_defaults())
    .build("pp-doclayoutv2.onnx")?;

let image = load_image("page.jpg")?;
let results = predictor.predict(vec![image])?;

for element in &results.elements[0] {
    println!("Type: {}, Score: {:.2}", element.element_type, element.score);
}

Available Predictors