LMSupply.Ocr

LMSupply.Ocr

A simple .NET library for local OCR (Optical Character Recognition) with automatic model downloading from HuggingFace. Features a 2-stage detection + recognition pipeline using PaddleOCR ONNX models.

Features

• 2-Stage Pipeline: Text detection (DBNet) followed by text recognition (CRNN with CTC decoding)
• Multi-language Support: 11 script recognizers covering 50+ language codes, including English, Korean, Chinese, Japanese, Arabic, Cyrillic and most Latin-script languages
• Automatic Model Download: Models are downloaded on-demand from HuggingFace (~10MB default)
• GPU Acceleration: Supports CUDA, DirectML, and CoreML
• Pure C# Implementation: No Python dependencies or external processes
• Polygon Support: Precise text region boundaries for rotated or curved text

Quick Start

using LMSupply.Ocr;

// Load default OCR pipeline (English)
await using var ocr = await LocalOcr.LoadAsync();

// Recognize text in an image
var result = await ocr.RecognizeAsync("document.png");

// Get all text
Console.WriteLine(result.FullText);

// Access individual text regions
foreach (var region in result.Regions)
{
    Console.WriteLine($"[{region.Confidence:P0}] {region.Text}");
    Console.WriteLine($"  Location: {region.BoundingBox}");
}

Language-Specific OCR

// Load OCR for Korean text
await using var ocr = await LocalOcr.LoadForLanguageAsync("ko");

// Or specify the recognition model explicitly
await using var ocr = await LocalOcr.LoadAsync(
    detectionModel: "default",
    recognitionModel: "crnn-korean-v3");

Supported Languages

Recognition is per script: each model reads one script's alphabet, and the caller picks the language. There is no script or language detection step — the models have no identification output to detect with.

Ask the library which codes it can resolve, rather than keeping your own list:

IEnumerable<string> codes = LocalOcr.GetSupportedLanguages();   // "en", "ko", "ja", "tr", ...

Region subtags are ignored ("tr-TR" resolves as "tr"). A code with no model falls back to the English recognizer and raises a trace warning naming the code — the English model cannot read another script, so its output for such text would otherwise look like a result.

Configuration Options

var options = new OcrOptions
{
    LanguageHint = "en",           // Language hint for auto model selection
    DetectionThreshold = 0.5f,      // Minimum detection confidence
    RecognitionThreshold = 0.5f,    // Regions below this recognition confidence are dropped (0 keeps all)
    BinarizationThreshold = 0.3f,   // DBNet binarization threshold
    UnclipRatio = 1.5f,             // Polygon expansion ratio
    UsePolygon = true,              // Use polygon coordinates
    Provider = ExecutionProvider.Auto,  // GPU acceleration
    CacheDirectory = null,          // Custom cache directory
    DisableAutoDownload = false     // true: load only from the cache, throw if a file is missing
};

await using var ocr = await LocalOcr.LoadAsync(options: options);

Checking the Cache Before Loading

Loading a language downloads whatever it needs that is not cached. To ask first — before a download the user has not agreed to — check the language's cache status. It makes no network request:

var status = LocalOcr.GetCacheStatusForLanguage("ko");
if (!status.IsCached)
{
    foreach (var file in status.Missing)
        Console.WriteLine($"Would download {file.RepoId}/{file.Subfolder}/{file.FileName}");
}

The files checked are the ones LoadForLanguageAsync loads: the detection model plus the language's recognizer and dictionary. Languages of one script share a recognizer, so having English cached says nothing about Korean. DisableAutoDownload = true makes a load throw instead of downloading.

Detection Only

// Get text regions without recognition
var regions = await ocr.DetectAsync("document.png");

foreach (var region in regions)
{
    Console.WriteLine($"Found text at: {region.BoundingBox}");
}

Layout-Aware Text Extraction

var result = await ocr.RecognizeAsync("document.png");

// Get text with layout preserved (same-line regions joined with spaces)
var layoutText = result.GetTextWithLayout(lineTolerancePixels: 10);

Model Information

The library uses PaddleOCR ONNX models from HuggingFace:

• Repository: monkt/paddleocr-onnx
• Detection model: DBNet (~2.3MB)
• Recognition models: CRNN (~7-13MB depending on language)

Models are cached locally after first download.