OpenCV find contours of close objects - c++

Question

I am writing a C++ Application using the OpenCV library to detect objects in images. These images look like this:
http://fs1.directupload.net/images/150311/my6uczfn.png
The upper part of the image, which is black, can be ignored.
I know, that every pixel, which is not part of a desired object, will be colored in white. What I am trying to do is to find out how many objects of interest are on an image and where they are.
Up until now I wrote the following code:
Mat image = imread("2.png", CV_LOAD_IMAGE_COLOR);
if(!image.data)
{
std::cout << "Could not open or find the image." << std::endl;
}
Range range_rows(0, image.size().height);
Range range_columns_left(0, image.size().width);
Range range_columns_middle(image.size().width, image.size().width * 2);
Range range_columns_right(image.size().width * 2, image.size().width * 3);
Mat display_mat(image.size().height, image.size().width * 3, CV_8UC3);
Mat left(display_mat, range_rows, range_columns_left);
image.copyTo(left);
Mat classified_image;
threshold(image, classified_image, 254, 255, THRESH_BINARY);
Mat middle(display_mat, range_rows, range_columns_middle);
classified_image.copyTo(middle);
Mat cimage = Mat::zeros(image.size(), CV_8UC3);
Mat classified_grayscale_image;
cvtColor(classified_image, classified_grayscale_image, CV_RGB2GRAY);
std::vector< std::vector<cv::Point> > contours;
findContours(classified_grayscale_image, contours, CV_RETR_LIST, CV_CHAIN_APPROX_NONE);
for(size_t counter = 0; counter < contours.size(); counter++)
{
std::cout << "Contours size: " << contours[counter].size() << std::endl;
if(contours[counter].size() < 6)
continue;
Mat pointsf;
Mat(contours[counter]).convertTo(pointsf, CV_32F);
RotatedRect box = fitEllipse(pointsf);
drawContours(cimage, contours, (int)counter, Scalar::all(255), 1, 8);
ellipse(cimage, box, Scalar(0,0,255), 1, CV_AA);
std::cout << "Ellipse Parameter:\t";
ellipse(cimage, box.center, box.size*0.5f, box.angle, 0, 360, Scalar(0,255,255), 1, CV_AA);
Point2f vtx[4];
box.points(vtx);
for( int j = 0; j < 4; j++ )
line(cimage, vtx[j], vtx[(j+1)%4], Scalar(0,255,0), 1, CV_AA);
}
Mat right(display_mat, range_rows, range_columns_right);
cimage.copyTo(right);
namedWindow("Results", CV_WINDOW_AUTOSIZE);
imshow("Results", display_mat);
waitKey(0);
return 0;
The result looks like this:
http://fs1.directupload.net/images/150311/toiy3aes.png
As you see, the classification, what is an object and what is not, is not perfect, so 2 objects are recognized as one. The classification will be improved, but something like this can happen, if those objects are very close. Even more of a problem is, when they are touching each other.
How can I do a proper object recognition in the case shown above? Any ideas?

Answer 1

You have got few options:
use some filtering/thresholding method on your result image to split objects from each other. Otsu binarization should be enough in this case, alternatively you can try to use dilate operation.
inverse your result image and than use distance transform and Otsu binarization (or some other kind o thresholding - most of them should work fine). It will make you objects smaller but will make counting them much easier.
if you need to mark the objects as precise as possible you need to use more complicated method. Here there is an tutorial which use techniques i've described above and connected components and watershed.