HoughCircles Doesn't detect circles correctly in OpenCV

Question

I am using Visual Studio 2015, OpenCV.3 and EmguCV.3. My code represented below and the result is shown in the picture. I know that the problem is the input values of HoughCircles function, but I don't know which inputs are suitable for this picture. I appreciate any help.

                Image<Gray, byte> OriginalImage = new Image<Gray, byte>(Openfile.FileName);
                Image<Gray, byte> ResizedImage = OriginalImage.Resize(OriginalImage.Width / 2, OriginalImage.Height / 2, Emgu.CV.CvEnum.Inter.Cubic);

                //********** Convert Image to Binary
                Image<Gray, byte> smoothImg = 
                ResizedImage.SmoothGaussian(5);
                smoothImg._Erode(5);
                smoothImg._Dilate(5);
                Image<Gray, byte> BinaryImage = 
                smoothImg.ThresholdBinary(new Gray(20), new Gray(255));

                //********** Find Circles
                Image<Rgb, byte> ROIImgScaledCircles = ROIImgScaled.Convert<Rgb, byte>();
                CircleF[] circles = smoothImg.HoughCircles(
                    new Gray(180),//cannyThreshold
                    new Gray(60),//circleAccumulatorThreshold
                    2.0, //dp:Resolution of the accumulator used to detect centers of the circles
                    10.0, //min distance 
                    10, //min radius
                    128 //max radius
                    )[0]; //Get the circles from the first channel
                foreach (CircleF cir in circles)
                {
                    ROIImgScaledCircles.Draw(cir, new Rgb(235, 20, 30), 1);
                }                   
                pbxCircles.Image = ROIImgScaledCircles.ToBitmap();

Original Image:

Founded Circles:

Answer 1

Working with complete shapes, you may find it easier to work detect edges and then find contours. Here's an example:

    Image<Bgr, byte> original = new Image<Bgr, byte>(@"E:\Downloads\original.jpg");
    UMat grayscale = new UMat();            
    UMat pyrdown = new UMat();
    UMat canny = new UMat();

    double cannyThreshold = 128;

    CvInvoke.CvtColor(original, grayscale, ColorConversion.Bgr2Gray);        
    // remove noise and run edge detection
    CvInvoke.PyrDown(grayscale, pyrdown);
    CvInvoke.PyrUp(pyrdown, grayscale);
    CvInvoke.Canny(grayscale, canny, cannyThreshold, cannyThreshold * 2);

    Image<Bgr, byte> result = original.Copy();
    // find and draw circles   
    VectorOfVectorOfPoint contours = new VectorOfVectorOfPoint();
    CvInvoke.FindContours(canny, contours, null, RetrType.List, ChainApproxMethod.ChainApproxSimple);
    //CvInvoke.DrawContours(result, contours, -1, new MCvScalar(0, 0, 255));
    for (int i = 0; i < contours.Size; i++)
    {
        Ellipse ellipse = new Ellipse(CvInvoke.FitEllipse(contours[i]));
        result.Draw(ellipse, new Bgr(Color.Red), 1);
    }

    result.Save(@"E:\Downloads\circles.jpg");

Here's the result, from left to right:

1. The original image
2. The blurred image (using pyrdown/pyrup)
3. Result from canny edge detection
4. Reconstructed circles from contours

Answer 2

Here is a solution (based on OpenCvSharp, not on emgucv, which allows C# code to be very close to all OpenCV code that you can find in C++ or Python, but you can easily convert it back to emgucv).

I've removed the Erode and Dilate step (which in this case just destroy the original image too much).

What I used is a loop on hough circle calls (varying the inverse ratio to accumulator resolution) to ensure that I detect more than one circle, and not the circles I'm not interested in.

  int blurSize = 5;
  using (var src = new Mat("2Okrv.jpg"))
  using (var gray = src.CvtColor(ColorConversionCodes.BGR2GRAY))
  using (var blur = gray.GaussianBlur(new Size(blurSize, blurSize), 0))
  using (var dst = src.Clone())
  {
      // this hashset will automatically store all "unique" detected circles
      // circles are stored modulo some "espilon" value, set to 5 here (half of min size of hough circles below)
      var allCircles = new HashSet<CircleSegment>(new CircleEqualityComparer { Epsilon = 5 });

      // vary inverse ratio of accumulator resolution
      // depending on image, you may vary start/end/step
      for (double dp = 1; dp < 5; dp += 0.2)
      {
          // we use min dist = 1, to make sure we can detect concentric circles
          // we use standard values for other parameters (canny, ...)
          // we use your min max values (the max may be important when dp varies)
          var circles = Cv2.HoughCircles(blur, HoughMethods.Gradient, dp, 1, 100, 100, 10, 128);
          foreach (var circle in circles)
          {
              allCircles.Add(circle);
          }
      }

      // draw final list of unique circles
      foreach (var circle in allCircles)
      {
          Cv2.Circle(dst, circle.Center, (int)circle.Radius, Scalar.FromRgb(235, 20, 30), 1);
      }

      // display images
      using (new Window("src image", src))
      using (new Window("dst image", dst))
      {
          Cv2.WaitKey();
      }
  }

  public class CircleEqualityComparer : IEqualityComparer<CircleSegment>
  {
      public double Epsilon { get; set; }

      public bool Equals(CircleSegment x, CircleSegment y) => x.Center.DistanceTo(y.Center) <= Epsilon && Math.Abs(x.Radius - y.Radius) <= Epsilon;

      // bit of a hack... we return a constant so only Equals is used to compare two circles
      // since we have only few circles that's ok, we don't play with millions...
      public int GetHashCode(CircleSegment obj) => 0;
  }

Here is the result: