/// <summary>
/// Recognized single char
/// </summary>
/// <param name="cleanLetter">B&W thresholded image with single char as white on black background</param>
/// <returns>Best match recognition data</returns>
RecognizedLetter ProcessSingleLetter(Mat cleanLetter)
{
// ocr
int[] classes;
double[] confidences;
ocr.Eval(cleanLetter, out classes, out confidences);
// process
if (confidences.Length > 0)
{
var unit = new RecognizedLetter();
unit.Data = vocabulary[classes[0]].ToString();
unit.Confidence = confidences[0];
return(unit);
}
return(null);
}
/// <summary>
/// Recognizes letter on the image
/// </summary>
/// <param name="clean">B&W thresholded image with black background and white letters</param>
/// <param name="possibilities">ROI as a list of RotatedRect's with letters marked</param>
/// <returns>Recognized letters</returns>
IList <RecognizedLetter> RecognizeLetters(Mat clean, IList <RotatedRect> possibilities)
{
var result = new List <RecognizedLetter>();
foreach (RotatedRect potentialLetter in possibilities)
{
var corners = Cv2.BoxPoints(potentialLetter);
var chunk = clean.UnwrapShape(corners, 64);
// check aspect as scanner might have issues with some "tall" chars like { I, J, L, etc. } and we can
// help the detection here: as there are no that much wide chars, we presume all strongly disproportional chars
// to be tall and reduce rotation options
Mat[] letters = null;
var size = potentialLetter.Size;
var aspect = Math.Max(size.Width, size.Height) / Math.Min(size.Width, size.Height);
if (aspect >= 2.0)
{
if (chunk.Cols > chunk.Rows) // it lays on "broadside", let's make it tall again
{
chunk = chunk.Rotate(RotateFlags.Rotate90Clockwise);
}
letters = new Mat[]
{
chunk,
chunk.Rotate(RotateFlags.Rotate180)
};
}
else
{
// as we're not sure how exactly is letter rotated, we recognize it 4 times,
// with a 90 degrees step, than the best match (best "confidence" from OCR) is
// used
letters = new Mat[]
{
chunk,
chunk.Rotate(RotateFlags.Rotate90Clockwise),
chunk.Rotate(RotateFlags.Rotate180),
chunk.Rotate(RotateFlags.Rotate90CounterClockwise)
};
}
// ocr
int selectedIndex = -1;
RecognizedLetter unit = null;
var lowerCases = new List <KeyValuePair <int, RecognizedLetter> >();
for (int i = 0; i < letters.Length; ++i)
{
var local = ProcessSingleLetter(letters[i]);
if (null != local && (null == unit || (null != unit && null != local && local.Confidence > unit.Confidence)))
{
if (char.IsUpper(local.Data[0]))
{
unit = local;
selectedIndex = i;
}
else
{
lowerCases.Add(new KeyValuePair <int, RecognizedLetter>(i, local));
}
}
//Cv2.ImShow(string.Format("#{0}: {1} - {2:00.0}%", i, local.Data, local.Confidence * 100), letters[i]);
}
if (null == unit && lowerCases.Count > 0)
{
lowerCases.Sort((x, y) => x.Value.Confidence.CompareTo(y.Value.Confidence));
unit = lowerCases[0].Value;
selectedIndex = lowerCases[0].Key;
}
// process
if (null != unit)
{
// this will show each separated letter in it's own window, un-wrapped and rotated according
// to the best found match
//Cv2.ImShow(unit.Data, letters[selectedIndex]);
unit.Rect = Array.ConvertAll(corners, p => new Point(Math.Round(p.X), Math.Round(p.Y)));
unit.Angle = (potentialLetter.Angle + selectedIndex * 90) % 360;
result.Add(unit);
}
}
return(result);
}