I have hit a roadblock with this issue that I have in my project.
i have an image with three printed objects and i want to check if the objects moves up or down during printing. If it does above/beyond the threshold limit it needs to be rejected.
E.G:
The above image shows a good sample, i select an ROI around all the 3 objects to reduce inspection time. But the problem is, the product shifts a bit during printing, although the print is good, the image apprears shifted within my ROI and the coordinates changes which leads them to get rejected.
EG:
How do I keep the image centered(ROI'd image) with reference to the center object (center object never shifts during printing).
i'm using template matching to find the objects as follows :
Image<Bgr, Byte> templateImage = img_tofind;
Image<Bgr, Byte> sourceImage = ROI_Image;
using (Image<Gray, float> imgMatch =
sourceImage.MatchTemplate(templateImage,
Emgu.CV.CvEnum.TM_TYPE.CV_TM_CCOEFF_NORMED))
{
Point[] MAX_Loc, Min_Loc;
double[] min, max;
imgMatch.MinMax(out min, out max, out Min_Loc, out
MAX_Loc);
using (Image<Gray, double> RG_Image =
imgMatch.Convert<Gray, double>().Copy())
{
if (max[0] > 0.60)
{
Rectangle match = new Rectangle(MAX_Loc[0],
templateImage.Size);
sourceImage.Draw(match, new Bgr(Color.Red), 1);
int hw = match.Width;
int hh = match.Height;
int htlx = match.X;
int htly = match.Y;
hres1 = htlx + hw / 2; Object COG X
hres2 = htly + hh / 2; Object COG Y
}
}
ibresult.Image = sourceImage;
}
I want the ROI image to always be centered with reference to the center object.
Related
I am having three image of pan card for testing skew of image using emgucv and c#.
1st image which is on top Detected 180 degree working properly.
2nd image which is in middle Detected 90 dgree should detected as 180 degree.
3rd image Detected 180 degree should detected as 90 degree.
One observation I am having that i wanted to share here is when i crop unwanted part of image from up and down side of pan card using paint brush, it gives me expected result using below mention code.
Now i wanted to understand how i can remove the unwanted part using programming.
I have played with contour and roi but I am not able to figure out how to fit the same. I am not able to understand whether emgucv itself selects contour or I have to do something.
Please suggest any suitable code example.
Please check code below for angle detection and please help me. Thanks in advance.
imgInput = new Image<Bgr, byte>(impath);
Image<Gray, Byte> img2 = imgInput.Convert<Gray, Byte>();
Bitmap imgs;
Image<Gray, byte> imgout = imgInput.Convert<Gray, byte>().Not().ThresholdBinary(new Gray(50), new Gray(125));
VectorOfVectorOfPoint contours = new VectorOfVectorOfPoint();
Emgu.CV.Mat hier = new Emgu.CV.Mat();
var blurredImage = imgInput.SmoothGaussian(5, 5, 0 , 0);
CvInvoke.AdaptiveThreshold(imgout, imgout, 255, Emgu.CV.CvEnum.AdaptiveThresholdType.GaussianC, Emgu.CV.CvEnum.ThresholdType.Binary, 5, 45);
CvInvoke.FindContours(imgout, contours, hier, Emgu.CV.CvEnum.RetrType.External, Emgu.CV.CvEnum.ChainApproxMethod.ChainApproxSimple);
if (contours.Size >= 1)
{
for (int i = 0; i <= contours.Size; i++)
{
Rectangle rect = CvInvoke.BoundingRectangle(contours[i]);
RotatedRect box = CvInvoke.MinAreaRect(contours[i]);
PointF[] Vertices = box.GetVertices();
PointF point = box.Center;
PointF edge1 = new PointF(Vertices[1].X - Vertices[0].X, Vertices[1].Y - Vertices[0].Y);
PointF edge2 = new PointF(Vertices[2].X - Vertices[1].X, Vertices[2].Y - Vertices[1].Y);
double r = edge1.X + edge1.Y;
double edge1Magnitude = Math.Sqrt(Math.Pow(edge1.X, 2) + Math.Pow(edge1.Y, 2));
double edge2Magnitude = Math.Sqrt(Math.Pow(edge2.X, 2) + Math.Pow(edge2.Y, 2));
PointF primaryEdge = edge1Magnitude > edge2Magnitude ? edge1 : edge2;
double primaryMagnitude = edge1Magnitude > edge2Magnitude ? edge1Magnitude : edge2Magnitude;
PointF reference = new PointF(1, 0);
double refMagnitude = 1;
double thetaRads = Math.Acos(((primaryEdge.X * reference.X) + (primaryEdge.Y * reference.Y)) / (primaryMagnitude * refMagnitude));
double thetaDeg = thetaRads * 180 / Math.PI;
imgInput = imgInput.Rotate(thetaDeg, new Bgr());
imgout = imgout.Rotate(box.Angle, new Gray());
Bitmap bmp = imgout.Bitmap;
break;
}
}
The Problem
Let us start with the problem before the solution:
Your Code
When you submit code, asking for help, at least make some effort to "clean" it. Help people help you! There's so many lines of code here that do nothing. You declare variables that are never used. Add some comments that let people know what it is that you think your code should do.
Bitmap imgs;
var blurredImage = imgInput.SmoothGaussian(5, 5, 0, 0);
Rectangle rect = CvInvoke.BoundingRectangle(contours[i]);
PointF point = box.Center;
double r = edge1.X + edge1.Y;
// Etc
Adaptive Thresholding
The following line of code produces the following images:
CvInvoke.AdaptiveThreshold(imgout, imgout, 255, Emgu.CV.CvEnum.AdaptiveThresholdType.GaussianC, Emgu.CV.CvEnum.ThresholdType.Binary, 5, 45);
Image 1
Image 2
Image 3
Clearly this is not what you're aiming for since the primary contour, the card edge, is completely lost. As a tip, you can always use the following code to display images at runtime to help you with debugging.
CvInvoke.NamedWindow("Output");
CvInvoke.Imshow("Output", imgout);
CvInvoke.WaitKey();
The Soltuion
Since your in example images the card is primarily a similar Value (in the HSV sense) to the background. I do not think simple gray scale thresholding is the correct approach in this case. I purpose the following:
Algorithm
Use Canny Edge Detection to extract the edges in the image.
Dilate the edges so as the card content combines.
Use Contour Detection to filter for the combined edges with the largest bounding.
Fit this primary contour with a rotated rectangle in order to extract the corner points.
Use the corner points to define a transformation matrix to be applied using WarpAffine.
Warp and crop the image.
The Code
You may wish to experiment with the parameters of the Canny Detection and Dilation.
// Working Images
Image<Bgr, byte> imgInput = new Image<Bgr, byte>("Test1.jpg");
Image<Gray, byte> imgEdges = new Image<Gray, byte>(imgInput.Size);
Image<Gray, byte> imgDilatedEdges = new Image<Gray, byte>(imgInput.Size);
Image<Bgr, byte> imgOutput;
// 1. Edge Detection
CvInvoke.Canny(imgInput, imgEdges, 25, 80);
// 2. Dilation
CvInvoke.Dilate(
imgEdges,
imgDilatedEdges,
CvInvoke.GetStructuringElement(
ElementShape.Rectangle,
new Size(3, 3),
new Point(-1, -1)),
new Point(-1, -1),
5,
BorderType.Default,
new MCvScalar(0));
// 3. Contours Detection
VectorOfVectorOfPoint inputContours = new VectorOfVectorOfPoint();
Mat hierarchy = new Mat();
CvInvoke.FindContours(
imgDilatedEdges,
inputContours,
hierarchy,
RetrType.External,
ChainApproxMethod.ChainApproxSimple);
VectorOfPoint primaryContour = (from contour in inputContours.ToList()
orderby contour.GetArea() descending
select contour).FirstOrDefault();
// 4. Corner Point Extraction
RotatedRect bounding = CvInvoke.MinAreaRect(primaryContour);
PointF topLeft = (from point in bounding.GetVertices()
orderby Math.Sqrt(Math.Pow(point.X, 2) + Math.Pow(point.Y, 2))
select point).FirstOrDefault();
PointF topRight = (from point in bounding.GetVertices()
orderby Math.Sqrt(Math.Pow(imgInput.Width - point.X, 2) + Math.Pow(point.Y, 2))
select point).FirstOrDefault();
PointF botLeft = (from point in bounding.GetVertices()
orderby Math.Sqrt(Math.Pow(point.X, 2) + Math.Pow(imgInput.Height - point.Y, 2))
select point).FirstOrDefault();
PointF botRight = (from point in bounding.GetVertices()
orderby Math.Sqrt(Math.Pow(imgInput.Width - point.X, 2) + Math.Pow(imgInput.Height - point.Y, 2))
select point).FirstOrDefault();
double boundingWidth = Math.Sqrt(Math.Pow(topRight.X - topLeft.X, 2) + Math.Pow(topRight.Y - topLeft.Y, 2));
double boundingHeight = Math.Sqrt(Math.Pow(botLeft.X - topLeft.X, 2) + Math.Pow(botLeft.Y - topLeft.Y, 2));
bool isLandscape = boundingWidth > boundingHeight;
// 5. Define warp crieria as triangles
PointF[] srcTriangle = new PointF[3];
PointF[] dstTriangle = new PointF[3];
Rectangle ROI;
if (isLandscape)
{
srcTriangle[0] = botLeft;
srcTriangle[1] = topLeft;
srcTriangle[2] = topRight;
dstTriangle[0] = new PointF(0, (float)boundingHeight);
dstTriangle[1] = new PointF(0, 0);
dstTriangle[2] = new PointF((float)boundingWidth, 0);
ROI = new Rectangle(0, 0, (int)boundingWidth, (int)boundingHeight);
}
else
{
srcTriangle[0] = topLeft;
srcTriangle[1] = topRight;
srcTriangle[2] = botRight;
dstTriangle[0] = new PointF(0, (float)boundingWidth);
dstTriangle[1] = new PointF(0, 0);
dstTriangle[2] = new PointF((float)boundingHeight, 0);
ROI = new Rectangle(0, 0, (int)boundingHeight, (int)boundingWidth);
}
Mat warpMat = new Mat(2, 3, DepthType.Cv32F, 1);
warpMat = CvInvoke.GetAffineTransform(srcTriangle, dstTriangle);
// 6. Apply the warp and crop
CvInvoke.WarpAffine(imgInput, imgInput, warpMat, imgInput.Size);
imgOutput = imgInput.Copy(ROI);
imgOutput.Save("Output1.bmp");
Two extension methods are used:
static List<VectorOfPoint> ToList(this VectorOfVectorOfPoint vectorOfVectorOfPoint)
{
List<VectorOfPoint> result = new List<VectorOfPoint>();
for (int contour = 0; contour < vectorOfVectorOfPoint.Size; contour++)
{
result.Add(vectorOfVectorOfPoint[contour]);
}
return result;
}
static double GetArea(this VectorOfPoint contour)
{
RotatedRect bounding = CvInvoke.MinAreaRect(contour);
return bounding.Size.Width * bounding.Size.Height;
}
Outputs
Meta Example
I am working on a project where I need to identify dots from IR lasers on a surface. I use for that a camera with IR filter
Some input images:
There can be several dots, too. So I tried to sharpen this image from webcam and then use FindContours method of Emgu CV.
There is my code:
public static Image<Gray, byte> Sharpen(Image<Gray, byte> image, int w, int h, double sigma1, double sigma2, int k)
{
w = (w % 2 == 0) ? w - 1 : w;
h = (h % 2 == 0) ? h - 1 : h;
//apply gaussian smoothing using w, h and sigma
var gaussianSmooth = image.SmoothGaussian(w, h, sigma1, sigma2);
//obtain the mask by subtracting the gaussian smoothed image from the original one
var mask = image - gaussianSmooth;
//add a weighted value k to the obtained mask
mask *= k;
//sum with the original image
image += mask;
return image;
}
private void ProcessFrame(object sender, EventArgs arg)
{
Mat frame = new Mat();
if (_capture.Retrieve(frame, CameraDevice))
{
Image<Bgr, byte> original = frame.ToImage<Bgr, byte>();
Image<Gray, byte> img = Sharpen(frame.ToImage<Gray, byte>(), 100, 100, 100, 100, 30);
Image<Gray, byte> thresh = new Image<Gray, byte>(img.Size);
CvInvoke.PyrDown(img, thresh);
CvInvoke.PyrUp(thresh, thresh);
Image<Gray, byte> mask = new Image<Gray, byte>(thresh.Size);
Image<Gray, byte> cannyImg = thresh.Canny(10, 50);
VectorOfVectorOfPoint contours = new VectorOfVectorOfPoint();
Mat hierarchy = new Mat();
CvInvoke.FindContours(
cannyImg,
contours,
hierarchy,
RetrType.External,
ChainApproxMethod.ChainApproxSimple
);
Image<Bgr, byte> resultImage = img.Copy().Convert<Bgr, byte>();
int contCount = contours.Size;
for (int i = 0; i < contCount; i++)
{
using (VectorOfPoint contour = contours[i])
{
resultImage.Draw(CvInvoke.BoundingRectangle(contour), new Bgr(255, 0, 0), 5);
}
}
captureBox.Image = original.Bitmap;
cvBox.Image = resultImage.Bitmap;
}
}
Example of result image:
So it almost all the time works as I expect it to, but framerate is very low. I'm getting like 10-15 fps with resolution of 640x480. I need to be able to do the same thing for 1920x1080 with at least 30 fps. It's my first time with OpenCV and Emgu.CV. What can I do to make it perform better?
I solved this just setting the threshold, so that image turns black and white only. By adjusting the threshold I was able to achieve the same results if not better in terms of clarity, but also performance drastically improved since there is not heavy processing going on
Here is a snippet with ARCore library instead on EmguCV
var bitmap = eventArgs.Frame;
var filter = new Grayscale(0.2125, 0.7154, 0.0721);
var grayImage = filter.Apply(bitmap);
var thresholdFilter = new Threshold(CurrentThreshold);
thresholdFilter.ApplyInPlace(grayImage);
var blobCounter = new BlobCounter();
blobCounter.ProcessImage(grayImage);
var rectangles = blobCounter.GetObjectsRectangles();
I would like to do motion detection in C# (using EmguCV 3.0) to remove object in motion or in foreground to draw an overlay.
Here is a sample test I done with a Kinect (because It's a depth camera)
How can I get started with EmguCV 3.0 ?
I tried many background removal code that do not work
It seems OpticalFlow is a good start but there si no example in EmguCV 3.0
If I find the largest blob how can I find its contours ?
Can someone help me to get started ?
EDIT: 17/06/2015
In EmguCV3.0.0 RC I don't see OpticalFlow in the package and documentation:
http://www.emgu.com/wiki/files/3.0.0-rc1/document/html/b72c032d-59ae-c36f-5e00-12f8d621dfb8.htm
There is only : DenseOpticalFlow, OpticalFlowDualTVL1 ???
This is a AbsDiff Code:
var grayFrame = frame.Convert<Gray, Byte>();
var motionFrame = grayFrame.AbsDiff(backFrame)
.ThresholdBinary(new Gray(20), new Gray(255))
.Erode(2)
.Dilate(2);
Result:
I don't know how to get the motion in white ?
This is the Blob Code:
Image<Bgr, Byte> smoothedFrame = new Image<Bgr, byte>(frame.Size);
CvInvoke.GaussianBlur(frame, smoothedFrame, new Size(3, 3), 1); //filter out noises
Mat forgroundMask = new Mat();
fgDetector.Apply(smoothedFrame, forgroundMask);
CvBlobs blobs = new CvBlobs();
blobDetector.Detect(forgroundMask.ToImage<Gray, byte>(), blobs);
blobs.FilterByArea(400, int.MaxValue);
blobTracker.Update(blobs, 1.0, 0, 1);
foreach (var pair in blobs) {
CvBlob b = pair.Value;
CvInvoke.Rectangle(frame, b.BoundingBox, new MCvScalar(255.0, 255.0, 255.0), 2);
}
Result:
Why so much false positive ?
This is a MOG2 Code:
forgroundDetector.Apply(frame, forgroundMask);
motionHistory.Update(forgroundMask);
var motionMask = GetMotionMask();
Image<Bgr, Byte> motionImage = new Image<Bgr, byte>(motionMask.Size);
CvInvoke.InsertChannel(motionMask, motionImage, 0);
Rectangle[] rects;
using (VectorOfRect boundingRect = new VectorOfRect()) {
motionHistory.GetMotionComponents(segMask, boundingRect);
rects = boundingRect.ToArray();
}
foreach (Rectangle comp in rects) { ...
Result:
If I select the biggest Area how can I get the contour of the object ?
First, I can give you some example Optical Flow code.
Let oldImage and newImage be variables that hold the previous and current frame. In my code, it's of type Image<Gray, Byte>.
// prep containers for x and y vectors
Image<Gray, float> velx = new Image<Gray, float>(newImage.Size);
Image<Gray, float> vely = new Image<Gray, float>(newImage.Size);
// use the Horn and Schunck dense optical flow algorithm.
OpticalFlow.HS(oldImage, newImage, true, velx, vely, 0.1d, new MCvTermCriteria(100));
// color each pixel
Image<Hsv, Byte> coloredMotion = new Image<Hsv, Byte>(newImage.Size);
for (int i = 0; i < coloredMotion.Width; i++)
{
for (int j = 0; j < coloredMotion.Height; j++)
{
// Pull the relevant intensities from the velx and vely matrices
double velxHere = velx[j, i].Intensity;
double velyHere = vely[j, i].Intensity;
// Determine the color (i.e, the angle)
double degrees = Math.Atan(velyHere / velxHere) / Math.PI * 90 + 45;
if (velxHere < 0)
{
degrees += 90;
}
coloredMotion.Data[j, i, 0] = (Byte) degrees;
coloredMotion.Data[j, i, 1] = 255;
// Determine the intensity (i.e, the distance)
double intensity = Math.Sqrt(velxHere * velxHere + velyHere * velyHere) * 10;
coloredMotion.Data[j, i, 2] = (intensity > 255) ? 255 : intensity;
}
}
// coloredMotion is now an image that shows intensity of motion by lightness
// and direction by color.
Regarding the larger question of how to remove the foreground:
If I had a way to get a static background image, that's the best way to start. Then, the foreground would be detected by the AbsDiff method and using Erode and Dilate or Gaussian to smooth the image, then use blob detection.
For simple foreground detection, I found Optical Flow to be way too much processing (8fps max), whereas the AbsDiff method was just as accurate but had no effect on framerate.
Regarding contours, if you're merely looking to find the size, position, and other moments, then the blob detection in the AbsDiff tutorial above seems to be sufficient, which uses Image.FindContours(...).
If not, I would start looking at the CvBlobDetector class as used in this tutorial. There's a built-in DrawBlob function that might come in handy.
In reference to: How to detect and count a spiral's turns
I am not able to get count even in the pixel based calculation also.
If I have attached image how to start with the counting the turns.
I tried the FindContours(); but doesn't quite get the turns segregated which it can't. Also the matchshape() I have the similarity factor but for whole coil.
So I tried as follows for turn count:
public static int GetSpringTurnCount()
{
if (null == m_imageROIed)
return -1;
int imageWidth = m_imageROIed.Width;
int imageHeight = m_imageROIed.Height;
if ((imageWidth <= 0) || (imageHeight <= 0))
return 0;
int turnCount = 0;
Image<Gray, float> imgGrayF = new Image<Gray, float>(imageWidth, imageHeight);
CvInvoke.cvConvert(m_imageROIed, imgGrayF);
imgGrayF = imgGrayF.Laplace(1); // For saving integer overflow.
Image<Gray, byte> imgGray = new Image<Gray, byte>(imageWidth, imageHeight);
Image<Gray, byte> cannyEdges = new Image<Gray, byte>(imageWidth, imageHeight);
CvInvoke.cvConvert(imgGrayF, imgGray);
cannyEdges = imgGray.Copy();
//cannyEdges = cannyEdges.ThresholdBinary(new Gray(1), new Gray(255));// = cannyEdges > 0 ? 1 : 0;
cannyEdges = cannyEdges.Max(0);
cannyEdges /= 255;
Double[] sumRow = new Double[cannyEdges.Cols];
//int sumRowIndex = 0;
int Rows = cannyEdges.Rows;
int Cols = cannyEdges.Cols;
for (int X = 0; X < cannyEdges.Cols; X++)
{
Double sumB = 0;
for (int Y = 0; Y < cannyEdges.Rows; Y ++)
{
//LineSegment2D lines1 = new LineSegment2D(new System.Drawing.Point(X, 0), new System.Drawing.Point(X, Y));
Double pixels = cannyEdges[Y, X].Intensity;
sumB += pixels;
}
sumRow[X] = sumB;
}
Double avg = sumRow.Average();
List<int> turnCountList = new List<int>();
int cnt = 0;
foreach(int i in sumRow)
{
sumRow[cnt] /= avg;
if(sumRow[cnt]>3.0)
turnCountList.Add((int)sumRow[cnt]);
cnt++;
}
turnCount = turnCountList.Count();
cntSmooth = cntSmooth * 0.9f + (turnCount) * 0.1f;
return (int)cntSmooth;
}
I am next trying surf.
==================================================
Edit: Adding samples. If you like it do it.
==================================================
Edit: Tried another algo:
ROI then Rotate ( biggest thin light blue rectangle )
GetMoments() shrink ROI height and position.Y using the moment.
Set the shrinked ROI and ._And() it with a blank image. ( Gray region with green rectangle )
cut the image into half-half.
contour and fit ellipse.
get maximum number of fitted ellipses.
Later will work on better algos and results.
Assuming the bigger cluster of white colour is the spring
--EDIT--
Apply inverse threshold to the picture and fill corners with flood fill algorithm.
Find the rotated bounding box of the biggest white cluster using findContours and minAreaRect
Trace the box longer axis doing the following
for each pixel along the axis trace axis line perpendicular going through current pixel
This line will cross the spring in minimum two points.
Find the point with the bigger distane from the axis
This will create collection on points similar to sine function
Count the peaks or clusters of this collection this will get twice the number of loops.
All this assuming you don't have high noise in the picture.
Hi have already function solution but one issue:
// The screenshot will be stored in this bitmap.
Bitmap capture = new Bitmap(rec.Width, rec.Height, PixelFormat.Format24bppRgb);
using (Graphics g = Graphics.FromImage(capture))
{
g.CopyFromScreen(rec.Location, new System.Drawing.Point(0, 0), rec.Size);
}
MCvSURFParams surfParam = new MCvSURFParams(500, false);
SURFDetector surfDetector = new SURFDetector(surfParam);
// Template image
Image<Gray, Byte> modelImage = new Image<Gray, byte>("template.jpg");
// Extract features from the object image
ImageFeature[] modelFeatures = surfDetector.DetectFeatures(modelImage, null);
// Prepare current frame
Image<Gray, Byte> observedImage = new Image<Gray, byte>(capture);
ImageFeature[] imageFeatures = surfDetector.DetectFeatures(observedImage, null);
// Create a SURF Tracker using k-d Tree
Features2DTracker tracker = new Features2DTracker(modelFeatures);
Features2DTracker.MatchedImageFeature[] matchedFeatures = tracker.MatchFeature(imageFeatures, 2);
matchedFeatures = Features2DTracker.VoteForUniqueness(matchedFeatures, 0.8);
matchedFeatures = Features2DTracker.VoteForSizeAndOrientation(matchedFeatures, 1.5, 20);
HomographyMatrix homography = Features2DTracker.GetHomographyMatrixFromMatchedFeatures(matchedFeatures);
// Merge the object image and the observed image into one image for display
Image<Gray, Byte> res = modelImage.ConcateVertical(observedImage);
#region draw lines between the matched features
foreach (Features2DTracker.MatchedImageFeature matchedFeature in matchedFeatures)
{
PointF p = matchedFeature.ObservedFeature.KeyPoint.Point;
p.Y += modelImage.Height;
res.Draw(new LineSegment2DF(matchedFeature.SimilarFeatures[0].Feature.KeyPoint.Point, p), new Gray(0), 1);
}
#endregion
#region draw the project region on the image
if (homography != null)
{
// draw a rectangle along the projected model
Rectangle rect = modelImage.ROI;
PointF[] pts = new PointF[] {
new PointF(rect.Left, rect.Bottom),
new PointF(rect.Right, rect.Bottom),
new PointF(rect.Right, rect.Top),
new PointF(rect.Left, rect.Top)
};
homography.ProjectPoints(pts);
for (int i = 0; i < pts.Length; i++)
pts[i].Y += modelImage.Height;
res.DrawPolyline(Array.ConvertAll<PointF, Point>(pts, Point.Round), true, new Gray(255.0), 2);
}
#endregion
pictureBoxScreen.Image = res.ToBitmap();
the result is:
And my problem is that, function homography.ProjectPoints(pts);
Get only first occurrence of pattern (white rectangle in pic above)
How i can Project all occurrence of template, respectively how I can get occurrence of template rectangle in image
I face a problem similar to yours in my master thesis. Basically you have two options:
Use a clustering such as Hierarchical k-means or a point density one such as DBSCAN (it depends on two parameters but you can make it threshold free in bidimensional R^2 space)
Use a multiple robust model fitting estimation techniques such as JLinkage. In this more advanced technique you clusters points that share an homography instead of cluster points that close to each other in euclidean space.
Once you partition your matches in "clusters" you can estimate homographies between matches belonging to correspondant clusters.