PaddleSharp

To perform full OCR using local V5 models on Windows, install the following NuGet packages:

Sdcb.PaddleInference
Sdcb.PaddleOCR
Sdcb.PaddleOCR.Models.Local
Sdcb.PaddleInference.runtime.win64.mkl
OpenCvSharp4.runtime.win

Use LocalFullModels to select a model and PaddleOcrAll to execute the OCR process. You can configure rotation detection and 180-degree classification via the PaddleOcrAll constructor/initializer.

FullOcrModel model = LocalFullModels.ChineseV5;

// ... (image loading code) ...

using (PaddleOcrAll all = new PaddleOcrAll(model, PaddleDevice.Mkldnn())
{
    AllowRotateDetection = true, /* Allow detection of angled text */ 
    Enable180Classification = false, /* Allow detection of text rotated > 90 degrees */
})
{
    using (Mat src = Cv2.ImDecode(sampleImageData, ImreadModes.Color))
    {
        PaddleOcrResult result = all.Run(src);
        Console.WriteLine("Detected all texts: \n" + result.Text);
        foreach (PaddleOcrResultRegion region in result.Regions)
        {
            Console.WriteLine($"Text: {region.Text}, Score: {region.Score}, RectCenter: {region.Rect.Center}, RectSize:    {region.Rect.Size}, Angle: {region.Rect.Angle}");
        }
    }
}

You can accelerate GPU inference using TensorRT by using the .And() method on PaddleDevice.Gpu() and providing a path to a shape information text file.

Important Considerations:

• Model Specificity: Shape info files (**.txt) are bound to specific models. If you are using a complex setup like PaddleOcrAll, you must provide different shape info files for the Detection, Classification, and Recognition models respectively.
• Cache Generation: The first run with TensorRT will generate a cache in %AppData%\Sdcb.PaddleInference\TensorRtCache. This initial run can take approximately 100 seconds. Subsequent runs will be significantly faster.
• Troubleshooting: If you encounter issues (e.g., using the wrong shape info for a model), delete the %AppData%\Sdcb.PaddleInference\TensorRtCache folder to force a regeneration of the cache.

using PaddleOcrAll all = new(model,
   PaddleDevice.Gpu().And(PaddleDevice.TensorRt("det.txt")),
   PaddleDevice.Gpu().And(PaddleDevice.TensorRt("cls.txt")),
   PaddleDevice.Gpu().And(PaddleDevice.TensorRt("rec.txt")))
{
   Enable180Classification = true,
   AllowRotateDetection = true,
};

To use Sdcb.PaddleOCR in an ASP.NET Core application, register QueuedPaddleOcrAll as a Singleton in your service builder. This allows you to manage OCR requests through a queue, which is useful for controlling concurrency.

In your controller, inject the QueuedPaddleOcrAll instance and use the .Run() method to process images. Note that you will likely need to use OpenCvSharp (e.g., Cv2.ImDecode) to convert uploaded files into a format suitable for the OCR engine.

// 1. Register in Service Collection
builder.Services.AddSingleton(s =>
{
    Action<PaddleConfig> device = builder.Configuration["PaddleDevice"] == "GPU" ? PaddleDevice.Gpu() : PaddleDevice.Mkldnn();
    return new QueuedPaddleOcrAll(() => new PaddleOcrAll(LocalFullModels.ChineseV5, device)
    {
        Enable180Classification = true,
        AllowRotateDetection = true,
    }, consumerCount: 1);
});

// 2. Use in Controller
public class OcrController : Controller
{
    private readonly QueuedPaddleOcrAll _ocr;

    public OcrController(QueuedPaddleOcrAll ocr) { _ocr = ocr; }

    [Route("ocr")]
    public async Task<OcrResponse> Ocr(IFormFile file)
    {
        using MemoryStream ms = new();
        using Stream stream = file.OpenReadStream();
        stream.CopyTo(ms);
        using Mat src = Cv2.ImDecode(ms.ToArray(), ImreadModes.Color);
        double scale = 1;
        using Mat scaled = src.Resize(default, scale, scale);

        Stopwatch sw = Stopwatch.StartNew();
        string textResult = (await _ocr.Run(scaled)).Text;
        sw.Stop();

        return new OcrResponse(textResult, sw.ElapsedMilliseconds);
    }
}