Slow performance using CPU

[malv007]

Hi,

Thanks for this library! It is great, but I find the performance to be slow, in my tests it is taking around 5 minutes to do just one page. Is it expected to be this slow? I remember trying this project a year ago, and I didn't think it was this slow.

This is the code I am using, I have tried changing a few of the settings but same performance result:

      const string modelsFolderName = "data/models/v5";
      string detPath = Path.Combine(modelsFolderName, "ch_PP-OCRv5_mobile_det.onnx");
      string clsPath = Path.Combine(modelsFolderName, "ch_ppocr_mobile_v2.0_cls_infer.onnx");
      string recPath = Path.Combine(modelsFolderName, "latin_PP-OCRv5_rec_mobile_infer.onnx");
      string keysPath = Path.Combine(modelsFolderName, "ppocrv5_latin_dict.txt");

      var text = new StringBuilder();

      var ocrEngin = GetOcrEngine(detPath, clsPath, recPath, keysPath);

      await foreach (var image in PDFtoImage.Conversion.ToImagesAsync(pdfMemory, true, null, new PDFtoImage.RenderOptions() { Dpi = 300, Grayscale = true, AntiAliasing = PDFtoImage.PdfAntiAliasing.All }, cancellationToken))
      {
		  RapidOcrNet.OcrResult ocrResult = ocrEngin.Detect(image, new RapidOcrNet.RapidOcrOptions()
		  {
			Padding = 100, //The image is surrounded by a white frame to improve the recognition rate. When the text frame does not correctly frame all the text, increase this value.
			ImgResize = 0,
			BoxScoreThresh = 0.6f, // The greater the threshold for whether the text-testing box retains, the lower the recall rate. Range of Values:[0, 1]The default value is 0.5. Text box confidence threshold, when the text box does not correctly frame all the text, reduce this value
			BoxThresh = 0.3f, // The text part and background part of the image are partitioned thresholds. The larger the value, the smaller the text part. Range of Values:[0, 1]The default value is 0.3.
			UnClipRatio = 1.5f, // Control the size of the text detection box, the larger the value, the larger the detection box as a whole. Range of Values:[1.6, 2.0]The default value is 1.6.
			DoAngle = false, // Enable (1) / Disable (0) text orientation detection. Text orientation detection is only required when the image is inverted (rotated 90 to 270 degrees).
			MostAngle = false
		  });
		  foreach (var block in ocrResult.TextBlocks)
		  {
			text.Append(block.GetText());
			text.Append("--");
		  }
      }

It looks like most of the time is spent here:

            // step: dbNet getTextBoxes
            var textBoxes = _textDetector.GetTextBoxes(src, scale, boxScoreThresh, boxThresh, unClipRatio) ?? [];

And if I drill down further, it seems like the bottleneck might be SkiaSharp. Does that align with what you are seeing?

This is one of the documents I am trying to OCR:

And the original document:
sample_doc.pdf

Is there anything I can do to speed this up?

Thanks so much!

[malv007]

I used some logic you had in a previous version, and now it takes a couple of seconds as opposed to minutes.

        byte[] roiBitmapBytes = new byte[gray8.BytesSize];
        nint buffer = Marshal.AllocHGlobal(gray8.BytesSize);
        try
        {
          roiBitmap.PeekPixels().ReadPixels(gray8, buffer, gray8.RowBytes);
          Marshal.Copy(buffer, roiBitmapBytes, 0, gray8.BytesSize);
        }
        finally
        {
          Marshal.FreeHGlobal(buffer);
        }

But I also saw your note about PeekPixels() not being reliable, but the OCR process seems to be working well. In which situations PeekPixels() isn't reliable?

[BobLd]

@malv007 thanks a lot for the issue and investigating. I need to have a look, and understand why my change caused slowness. Need to add benchmarks too

Ill try to have a look over the weekend

[BobLd]

@malv007 can you share the file where this change happened?

[malv007]

Definitely! It is your code after all :)

TextDetector.txt

Look for the method:
private static double GetScore(SKPoint[] contours, SKBitmap fMapMat)

[BobLd]

@malv007 thanks again for the help here. Issue is fixed and available in v1.0.1