public List <string> ProcessBodyImage(bool normAny = false)
{
Image <Bgr, byte> baseImg = new Image <Bgr, byte>(image);
Normalization <Bgr> rgb_norm = new Normalization <Bgr>(baseImg);
var normalized = rgb_norm.Normalize();
baseImg = rgb_norm.Result;
//saveImage(baseImg, Path.GetFileName(imagePath)+"base");
Image <Gray, byte> grayImg = baseImg.Convert <Gray, byte>();
//saveImage(grayImg, Path.GetFileName(imagePath)+"gray");
if (normalized)
{
Normalization <Gray> gray_norm = new Normalization <Gray>(grayImg);
if (normAny)
{
gray_norm.NormalizeAny();
}
grayImg = gray_norm.Result;
//saveImage(grayImg, Path.GetFileName(imagePath) + "gray_norm");
}
Image <Gray, byte> canny = grayImg.Canny(175, 320);
//saveImage(canny, "canny");
//detecting bounding boxes
var aContours = new VectorOfVectorOfPoint();
var aHierarchy = new Mat();
CvInvoke.FindContours(canny, aContours, aHierarchy, Emgu.CV.CvEnum.RetrType.List, Emgu.CV.CvEnum.ChainApproxMethod.LinkRuns, new Point(0, 0));
List <Rectangle> boxes = new List <Rectangle>();
for (int i = 0; i < aContours.Size; i++)
{
var item = aContours[i];
List <Point> points = new List <Point>();
for (int j = 0; j < item.Size; j++)
{
var item2 = item[j];
points.Add(new Point(item2.X, item2.Y));
}
var x_query = from Point p in points select p.X;
int xmin = x_query.Min();
int xmax = x_query.Max();
var y_query = from Point p in points select p.Y;
int ymin = y_query.Min();
int ymax = y_query.Max();
Rectangle r = new Rectangle(xmin, ymin, xmax - xmin, ymax - ymin);
boxes.Add(r);
}
//saveImage(drawBoxesOnImage(canny.Bitmap, boxes), Path.GetFileName(imagePath)+"test");
List <Tuple <Rectangle, List <Rectangle> > > itemsToUnite = new List <Tuple <Rectangle, List <Rectangle> > >();
//check if boxes contact more than 70%, if yes - unite them
for (int i = 0; i < boxes.Count; i++)
{
//contacts = new List<Rectangle>();
List <Rectangle> unions = new List <Rectangle>();
for (int j = i + 1; j < boxes.Count; j++)
{
//if (i == j)
// continue;
var b1 = boxes[i];
var b2 = boxes[j];
int dif = 1; //contact differenct
//check up/down & left/right contact
bool hasContact = false;
if (Math.Abs(b1.Bottom - b2.Top) == dif)
{
Rectangle left = b1.Left < b2.Left ? b1 : b2;
Rectangle right = b1.Right > b2.Right ? b1 : b2;
if (left.Right < right.Left)
{
continue;
}
hasContact = true;
}
else if (Math.Abs(b1.Right - b2.Right) == dif)
{
Rectangle top = b1.Top < b2.Top ? b1 : b2;
Rectangle bottom = b1.Bottom > b2.Bottom ? b1 : b2;
if (top.Bottom < bottom.Top)
{
continue;
}
hasContact = true;
}
if (hasContact)
{
//contacts.Add(b1);
//contacts.Add(b2);
//check if contact area if more than 70%
var length1 = b1.Right - b1.Left;
var length2 = b2.Right - b1.Left;
var length = Math.Max(b1.Right, b2.Right) - Math.Min(b1.Left, b2.Left);
if (length > 0)
{
var left_offset = Math.Max(b1.Left, b2.Left) - Math.Min(b1.Left, b2.Left);
var right_offset = Math.Max(b1.Right, b2.Right) - Math.Min(b1.Right, b2.Right);
var intersection = length - left_offset - right_offset;
var perc = 100 * intersection / (float)length;
if (perc >= 70)
{
unions.Add(b2);
}
}
}
}
//if (contacts.Any())
// saveImage(drawBoxesOnImage(canny.Bitmap, contacts), "contact_" + i);
//if (unions.Any())
itemsToUnite.Add(new Tuple <Rectangle, List <Rectangle> >(boxes[i], unions));
//if (contacts.Any())
// break;
}
//saveImage(drawBoxesOnImage(canny.Bitmap, contacts), "contact");
List <Rectangle> newBoxes = new List <Rectangle>();
foreach (var item in itemsToUnite)
{
if (item.Item2.Any())
{
var lst = item.Item2;
lst.Add(item.Item1);
Rectangle r = getBoundingBox(lst);
newBoxes.Add(r);
}
else
{
bool canAdd = true;
foreach (var i in itemsToUnite)
{
if (i.Item2.Contains(item.Item1))
{
canAdd = false;
break;
}
}
if (canAdd)
{
newBoxes.Add(item.Item1);
}
}
}
boxes = newBoxes;
//saveImage(drawBoxesOnImage(canny.Bitmap, boxes), Path.GetFileName(imagePath) + "unions");
//filter bounding boxes
float minHeight = 5;
boxes.RemoveAll(x => x.Height < minHeight);
boxes.RemoveAll(x => x.Height < x.Width);
boxes.RemoveAll(x => x.Height > canny.Height / 2);
boxes.RemoveAll(x => x.Width < 2);
//saveImage(drawBoxesOnImage(canny.Bitmap, boxes), Path.GetFileName(imagePath) + "filtered");
//detecting numbers bounding boxes
List <Rectangle> sums = new List <Rectangle>();
List <Rectangle> lefts = new List <Rectangle>();
List <Rectangle> rights = new List <Rectangle>();
List <Rectangle> extended = new List <Rectangle>();
boxes = boxes.OrderBy(x => x.X).ToList();
for (int i = 0; i < boxes.Count; i++)
{
var box = boxes[i];
int offsetWidth = (int)(box.Width / 3);
Rectangle offset1 = new Rectangle(box.X - offsetWidth, box.Y, offsetWidth, box.Height),
offset2 = new Rectangle(box.X + box.Width, box.Y, offsetWidth, box.Height);
Rectangle uni = Rectangle.Union(box, offset1);
uni = Rectangle.Union(uni, offset2);
extended.Add(uni);
lefts.Add(offset1);
rights.Add(offset2);
}
//saveImage(drawBoxesOnImage(canny.Bitmap, new Color[] { Color.Red, Color.Green, Color.Blue }, boxes, lefts, rights), "offsets");
//saveImage(drawBoxesOnImage(canny.Bitmap, extended), Path.GetFileName(imagePath) + "extended");
List <IntersectionHierarchyItem> intersections = new List <IntersectionHierarchyItem>();
foreach (var box in extended)
{
intersections.Add(findIntersectingHierarchy(extended, box));
}
List <Rectangle> result = new List <Rectangle>();
foreach (var box in intersections)
{
if (box.HasIntersection)
{
result.Add(box.Union);
}
}
result = result.Distinct().ToList();
//filtering horizontal rectangles
result.RemoveAll(x => x.Width <= x.Height);
//filtering rectangles by aspect ratio
result.RemoveAll(x =>
{
float aspectRatio = (float)x.Width / (float)x.Height;
return(aspectRatio > 0.75 && aspectRatio < 1.3);
});
//saveImage(drawBoxesOnImage(canny.Bitmap, result), Path.GetFileName(imagePath) + "filtered");
if (!result.Any())
{
if (!normAny)
{
return(ProcessBodyImage(true));
}
return(new List <string>());
}
List <Rectangle> bounding = new List <Rectangle>();
List <Rectangle[]> sRects = new List <Rectangle[]>();
List <List <string> > digitVariants = new List <List <string> >();
//cutting numbers from images
for (int j = 0; j < result.Count; j++)
{
var area = result[j];
//find source bounding boxes that are inside intersecting area
List <Rectangle> rects = findInnerRectangles(boxes, area);
//save(drawBoxesOnImage(canny, rects), imgNumber, "inner1_"+j);
//remove rectangles that are inside another rect
rects = removeInnerRectangles(rects);
//save(drawBoxesOnImage(canny, rects), imgNumber, "inner2_" + j);
//saveCoords(rects, imgNumber, "inner2_" + j);
//TODO: do intersection
rects = merge(rects);
sRects.Add(rects.ToArray());
bounding.Add(getBoundingBox(rects));
//saveImage(drawBoxesOnImage(canny.Bitmap, rects), "inner_" + j);
//saveCoords(rects, imgNumber, "inner_" + j);
//distinct list to prevent adding duplicating rectangles after merging
rects = rects.Distinct().ToList();
List <string> tesseractParts = new List <string>();
//cropping each rectangle and saving as image
if (digitsRecognitionMethod == DigitsRecognitionMethod.Tesseract || digitsRecognitionMethod == DigitsRecognitionMethod.Both)
{
List <string> digitVariant = new List <string>();
for (int i = 0; i < rects.Count; i++)
{
var gray = grayImg.Clone();
gray.ROI = rects[i];
Mat componentRoi = gray.Mat;
Mat thresholdedMat = gray.Mat;
CvInvoke.Threshold(componentRoi, thresholdedMat, 0, 255, Emgu.CV.CvEnum.ThresholdType.Otsu | Emgu.CV.CvEnum.ThresholdType.BinaryInv);
string digitLocation = FileManager.TempPng;
thresholdedMat.Save(digitLocation);
digitVariant.Add(digitLocation);
//save(thresholdedMat, imgNumber, "digit_" + j + "_" + i);
//save(crop(canny, rects[i]), imgNumber, "digit_" + j + "_" + i);
}
digitVariants.Add(digitVariant);
}
}
//saveImage(drawBoxesOnImage(canny.Bitmap, bounding), "bb");
List <string> numbersFinals = new List <string>();
if (digitsRecognitionMethod == DigitsRecognitionMethod.Tesseract || digitsRecognitionMethod == DigitsRecognitionMethod.Both)
{
foreach (var dvar in digitVariants)
{
string file = saveTesseract(dvar);
numbersFinals.Add(OCRParser.ParseTesseract(file));
}
}
if (digitsRecognitionMethod == DigitsRecognitionMethod.Neural || digitsRecognitionMethod == DigitsRecognitionMethod.Both)
{
//get max campatible bounding box
//var largestRect = bounding.Aggregate((r1, r2) => (((r1.Height * r1.Width) > (r2.Height * r2.Width)) || ()) ? r1 : r2);
int index = 0;
List <string> digitPaths = new List <string>();
if (bounding.Count > 0)
{
int maxArea = bounding[index].Height * bounding[index].Width;
int lastSubs = sRects[index].Length;
int goodAspects = checkGoodLetters(sRects[index]);
for (int i = 1; i < bounding.Count; i++)
{
//exclude elements that contain much more than 5 rectangles inside (this means that rectagles don't represent letters and numbers but other shapes)
int subs = sRects[i].Length;
if (subs > 5)
{
continue;
}
//exclude elements by aspect ratio
float aspectRatio = (float)bounding[i].Width / (float)bounding[i].Height;
const float MAX_ASPECT = 2.4f; //12 / 5
const float MIN_ASPECT = 1.7f;
//if (aspectRatio > MAX_ASPECT || aspectRatio < MIN_ASPECT)
// continue;
//if (lastSubs > subs)
// continue;
int area = bounding[i].Height * bounding[i].Width;
if (area > maxArea)
{
//check letters aspect ratio
int lets = checkGoodLetters(sRects[i]);
if (lets > goodAspects)
{
index = i;
maxArea = area;
lastSubs = subs;
goodAspects = lets;
}
}
}
//int index = bounding.IndexOf(largestRect);
var elems = sRects[index];
for (int i = 0; i < elems.Length; i++)
{
var gray = grayImg.Clone();
gray.ROI = elems[i];
Mat componentRoi = gray.Mat;
Mat thresholdedMat = gray.Mat;
CvInvoke.Threshold(componentRoi, thresholdedMat, 0, 255, Emgu.CV.CvEnum.ThresholdType.Otsu | Emgu.CV.CvEnum.ThresholdType.BinaryInv);
/*
* int s = (int)(0.05 * mat.Rows); // 5% of up-scaled size
* Mat elem = Cv2.GetStructuringElement(StructuringElementShape.Ellipse, new Size(2 * s + 1, 2 * s + 1), new Point(s, s));
* //Cv2.Erode(mat, mat, elem);
*/
int s = (int)(0.05 * thresholdedMat.Rows);
Mat elem = CvInvoke.GetStructuringElement(Emgu.CV.CvEnum.ElementShape.Ellipse, new Size(2 * s + 1, 2 * s + 1), new Point(s, s));
CvInvoke.Erode(thresholdedMat, thresholdedMat, elem, new Point(s, s), 1, Emgu.CV.CvEnum.BorderType.Reflect, default(MCvScalar));
string digitPath = FileManager.TempPng;
digitPaths.Add(digitPath);
thresholdedMat.Save(digitPath);
//save(thresholdedMat, imgNumber, "digit_" + "_" + i);
}
}
numbersFinals.Add(OCRParser.ParseNeural(digitPaths.ToArray()).Value);
}
return(numbersFinals);
}
public static string[] Recognize(IplImage input, TextDetectionParams _params, Chain[] chains, List <Tuple <Point2d, Point2d> > compBB, List <Tuple <CvPoint, CvPoint> > chainBB, DigitsRecognitionMethod digitsRecognition)
{
List <string> variants = new List <string>();
//convert to grayscale
IplImage grayImage = Cv.CreateImage(input.GetSize(), BitDepth.U8, 1);
Cv.CvtColor(input, grayImage, ColorConversion.RgbToGray);
for (int i = 0; i < chainBB.Count; i++)
{
Rect chainRect = new Rect(chainBB[i].Item1.X, chainBB[i].Item1.Y, chainBB[i].Item2.X - chainBB[i].Item1.X, chainBB[i].Item2.Y - chainBB[i].Item1.Y);
CvPoint center = new CvPoint((chainBB[i].Item1.X + chainBB[i].Item2.X) / 2, (chainBB[i].Item1.Y + chainBB[i].Item2.Y) / 2);
//work out if total width of chain is large enough
if (chainBB[i].Item2.X - chainBB[i].Item1.X < input.Width / _params.MaxImgWidthToTextRatio)
{
continue;
}
//eliminate chains with components of lower height than required minimum
int minHeight = chainBB[i].Item2.Y - chainBB[i].Item1.Y;
for (int j = 0; j < chains[i].components.Count; j++)
{
minHeight = Math.Min(minHeight, compBB[chains[i].components[j]].Item2.y - compBB[chains[i].components[j]].Item1.y);
}
if (minHeight < _params.MinCharacterHeight)
{
continue;
}
//invert direction if angle is in 3rd/4th quadrants
if (chains[i].direction.x < 0)
{
chains[i].direction.x = -chains[i].direction.x;
chains[i].direction.y = -chains[i].direction.y;
}
//work out chain angle
double theta_deg = 180 * Math.Atan2(chains[i].direction.y, chains[i].direction.x) / Math.PI;
if (Math.Abs(theta_deg) > _params.MaxAngle)
{
continue;
}
if ((chainBB.Count == 2) && (Math.Abs(theta_deg) > 5))
{
continue;
}
//Console.WriteLine("Chain #" + i + " angle: " + theta_deg + " degress");
//create copy of input image including only the selected components
Mat inputMat = new Mat(input);
Mat grayMat = new Mat(grayImage);
Mat componentsImg = Mat.Zeros(new Size(grayMat.Cols, grayMat.Rows), grayMat.Type());
//CvMat componentsImg = _componentsImg.ToCvMat();
Mat componentsImgRoi = null;
List <CvPoint> compCoords = new List <CvPoint>();
chains[i].components = chains[i].components.Distinct().ToList();
int order = 0;
//ordering components bounding boxes by x coord
var ordCompBB = compBB.OrderBy(x => x.Item1.x).ToList();
List <string> digits = new List <string>();
for (int j = 0; j < ordCompBB.Count; j++)
{
Rect roi = new Rect(ordCompBB[j].Item1.x, ordCompBB[j].Item1.y, ordCompBB[j].Item2.x - ordCompBB[j].Item1.x, ordCompBB[j].Item2.y - ordCompBB[j].Item1.y);
if (!chainRect.Contains(roi))
{
continue;
}
Mat componentRoi = new Mat(grayMat, roi);
compCoords.Add(new CvPoint(ordCompBB[j].Item1.x, ordCompBB[j].Item1.y));
compCoords.Add(new CvPoint(ordCompBB[j].Item2.x, ordCompBB[j].Item2.y));
compCoords.Add(new CvPoint(ordCompBB[j].Item1.x, ordCompBB[j].Item2.y));
compCoords.Add(new CvPoint(ordCompBB[j].Item2.x, ordCompBB[j].Item1.y));
Mat thresholded = new Mat(grayMat, roi);
Cv2.Threshold(componentRoi, thresholded, 0, 255, ThresholdType.Otsu | ThresholdType.BinaryInv);
componentsImgRoi = new Mat(componentsImg, roi);
Cv2.Threshold(componentRoi, componentsImgRoi, 0, 255, ThresholdType.Otsu | ThresholdType.BinaryInv);
//var size = thresholded.Size();
//digits.Add(new Bitmap(size.Width, size.Height, (int)thresholded.Step1(), System.Drawing.Imaging.PixelFormat.Format24bppRgb, thresholded.Data));
if (digitsRecognition == DigitsRecognitionMethod.Neural || digitsRecognition == DigitsRecognitionMethod.Both)
{
string file = FileManager.TempBitmap;
Cv2.ImWrite(file, thresholded);
try
{
digits.Add(file);
}
catch
{
GC.Collect();
GC.WaitForFullGCComplete();
}
//digits.Last().Save("test" + order + ".bmp");
order++;
}
//else if (digitsRecognition == DigitsRecognitionMethod.Tesseract || digitsRecognition == DigitsRecognitionMethod.Both)
//{
// DO NOTHING
//}
}
if (digitsRecognition == DigitsRecognitionMethod.Neural || digitsRecognition == DigitsRecognitionMethod.Both)
{
//TODO: neural recognition
var result = OCRParser.ParseNeural(digits.ToArray());
variants.Add(result.Value);
//variants.AddRange(OCRParser.ParseNeural(digits.ToArray()));
//variants.Add(BibOCR.OCRParser.ParseBib(digits.ToArray()));
}
if (digitsRecognition == DigitsRecognitionMethod.Tesseract || digitsRecognition == DigitsRecognitionMethod.Both)
{
CvRect _roi = GetBoundingBox(compCoords, new CvSize(input.Width, input.Height));
//ROI area can be null if outside of clipping area
if ((_roi.Width == 0) || (_roi.Height == 0))
{
continue;
}
//rotate each component coordinates
const int border = 3;
Mat _mat = new Mat(_roi.Height + 2 * border, _roi.Width + 2 * border, grayMat.Type());
Mat tmp = new Mat(grayMat, _roi);
//copy bounded box from rotated mat to new mat with borders - borders are needed to improve OCR success rate
Mat mat = new Mat(_mat, new Rect(border, border, _roi.Width, _roi.Height));
tmp.CopyTo(mat);
//resize image to improve OCR success rate
float upscale = 5.0f;
Cv2.Resize(mat, mat, new Size(0, 0), upscale, upscale);
//erode text to get rid of thin joints
int s = (int)(0.05 * mat.Rows); // 5% of up-scaled size
Mat elem = Cv2.GetStructuringElement(StructuringElementShape.Ellipse, new Size(2 * s + 1, 2 * s + 1), new Point(s, s));
//Cv2.Erode(mat, mat, elem);
//Cv2.Threshold(mat, mat, 0, 255, ThresholdType.Otsu | ThresholdType.BinaryInv);
string file = FileManager.TempPng;
Cv2.ImWrite(file, mat);
// TODO: Pass it to Tesseract API
variants.Add(OCRParser.ParseTesseract(file));
}
//for (int j = 0; j < digits.Count; j++)
// digits[j].Dispose();
digits.Clear();
GC.Collect();
GC.WaitForFullGCComplete(5000);
}
Cv.ReleaseImage(grayImage);
return(variants.Distinct().ToArray());
}