Starshot.OneOcr 0.2.0

Starshot.OneOcr

C# bindings for OneOCR — the OCR engine embedded in the Windows Photos app and Snipping Tool. Falls back to the system engine (Windows.Media.Ocr) when OneOCR is unavailable, and squashes the extra spaces that both engines insert between CJK characters.

This package does NOT contain the engine itself. OneOCR is a Microsoft proprietary component; you must obtain the two files yourself and place them next to your app exe:

• oneocr.dll (~39 MB)
• oneocr.onemodel (~56 MB, encrypted — the decryption key is included in this library)

Where to get them: copy from the local Windows Snipping Tool package (%ProgramFiles%\WindowsApps\Microsoft.ScreenSketch_*\SnippingTool\) or the Photos app package root, via Windows.Management.Deployment.PackageManager. Redistribution of these files is your own responsibility.

Requirements

• .NET 10, net10.0-windows10.0.26100.0 (Windows App SDK projection)
• Win2D 1.3.2+ (any compatible build)
• Serilog (used for engine logging)

Quick start

using Starshot.OneOcr;

// 1) On the UI thread (GPU readback must stay on the device's thread):
var (pixels, width, height, scale) = OcrHelper.PreparePixels(canvasBitmap);

// 2) Any thread — oneocr first, Windows.Media.Ocr fallback:
var result = await OcrHelper.RecognizeAsync(pixels, width, height, scale);

if (result is null)
{
    // no engine available at all (files missing AND no OCR language pack)
}
else if (result.Lines.Count == 0)
{
    // recognized nothing
}
else
{
    Console.WriteLine($"engine: {result.Engine}");  // OneOcr or Legacy (incl. fallback)
    foreach (var line in result.Lines)
    {
        Console.WriteLine(line.Text);          // full-line text, spaces already squashed
        foreach (var word in line.Words)
        {
            // word.Text + word.Rect (original-image pixel coordinates)
        }
    }
}

Pass preferOneOcr: false to RecognizeAsync to skip OneOCR and use the system engine directly.

API surface

Notes

• OneOCR failures (missing or corrupt engine files, runtime errors) automatically fall back to the system engine; check OcrResult.Engine to see which engine actually processed the image.
• The OneOCR pipeline is created lazily on first use and cached for the process lifetime; RunOcrPipeline calls are serialized with a lock (thread safety of the native engine is unknown).
• Both engines' line text is token-space-joined; SquashCjkSpaces removes spaces only when neither neighbor is a Latin letter/digit — CJK gets glued, English keeps its spaces.
• Coordinates from oversized inputs (legacy engine MaxImageDimension limit) are pre-scaled on the GPU and multiplied back so both engines return original-image pixels.