Detecting crosses in an image

Question

I am working on a program to detect the tips of a probing device and analyze the color change during probing. The input/output mechanisms are more or less in place. What I need now is the actual meat of the thing: detecting the tips.

In the images below, the tips are at the center of the crosses. I thought of applying BFS to the images after some threshold'ing but was then stuck and didn't know how to proceed. I then turned to OpenCV after reading that it offers feature detection in images. However, I am overwhelmed by the vast amount of concepts and techniques utilized here and again, clueless about how to proceed.

Am I looking at it the right way? Can you give me some pointers?

Image extracted from short video

Binary version with threshold set at 95

Answer 1

Template Matching Approach

Here is a simple matchTemplate solution, that is similar to the approach that Guy Sirton mentions.

Template matching will work as long as you don't have much scaling or rotation occurring with your target.

Here is the template that I used:

Here is the code I used to detect several of the unobstructed crosses:

#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char* argv[])
{
    string inputName = "crosses.jpg";
    string outputName = "crosses_detect.png";
    Mat img   = imread( inputName, 1);
    Mat templ = imread( "crosses-template.jpg", 1);

    int resultCols =  img.cols - templ.cols + 1;
    int resultRows = img.rows - templ.rows + 1;
    Mat result( resultCols, resultRows, CV_32FC1 );

    matchTemplate(img, templ, result, CV_TM_CCOEFF);
    normalize(result, result, 0, 255.0, NORM_MINMAX, CV_8UC1, Mat());

    Mat resultMask;
    threshold(result, resultMask, 180.0, 255.0, THRESH_BINARY);

    Mat temp = resultMask.clone();
    vector< vector<Point> > contours;
    findContours(temp, contours, CV_RETR_EXTERNAL, CV_CHAIN_APPROX_SIMPLE, Point(templ.cols / 2, templ.rows / 2));

    vector< vector<Point> >::iterator i;
    for(i = contours.begin(); i != contours.end(); i++)
    {
        Moments m = moments(*i, false);
        Point2f centroid(m.m10 / m.m00, m.m01 / m.m00);
        circle(img, centroid, 3, Scalar(0, 255, 0), 3);
    }

    imshow("img", img);
    imshow("results", result);
    imshow("resultMask", resultMask);

    imwrite(outputName, img);

    waitKey(0);

    return 0;
}

This results in this detection image:

This code basically sets a threshold to separate the cross peaks from the rest of the image, and then detects all of those contours. Finally, it computes the centroid of each contour to detect the center of the cross.

Shape Detection Alternative

Here is an alternative approach using triangle detection. It doesn't seems as accurate as the matchTemplate approach, but might be an alternative you could play with.

Using findContours we detect all the triangles in the image, which results in the following:

Then I noticed all the triangle vertices cluster near the cross center, so then these clusters are used to centroid the cross center point shown below:

Finally, here is the code that I used to do this:

#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <iostream>
#include <list>

using namespace cv;
using namespace std;

vector<Point> getAllTriangleVertices(Mat& img, const vector< vector<Point> >& contours);
double euclideanDist(Point a, Point b);

vector< vector<Point> > groupPointsWithinRadius(vector<Point>& points, double radius);
void printPointVector(const vector<Point>& points);
Point computeClusterAverage(const vector<Point>& cluster);

int main(int argc, char* argv[])
{
    Mat img   = imread("crosses.jpg", 1);
    double resizeFactor = 0.5;
    resize(img, img, Size(0, 0), resizeFactor, resizeFactor);

    Mat momentImg = img.clone();

    Mat gray;
    cvtColor(img, gray, CV_BGR2GRAY);

    adaptiveThreshold(gray, gray, 255.0, ADAPTIVE_THRESH_MEAN_C, THRESH_BINARY, 19, 15);
    imshow("threshold", gray);
    waitKey();

    vector< vector<Point> > contours;
    findContours(gray, contours, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE);

    vector<Point> allTriangleVertices = getAllTriangleVertices(img, contours);

    imshow("img", img);
    imwrite("shape_detect.jpg", img);
    waitKey();

    printPointVector(allTriangleVertices);
    vector< vector<Point> > clusters = groupPointsWithinRadius(allTriangleVertices, 10.0*resizeFactor);
    cout << "Number of clusters: " << clusters.size() << endl;

    vector< vector<Point> >::iterator cluster;
    for(cluster = clusters.begin(); cluster != clusters.end(); ++cluster)
    {
        printPointVector(*cluster);

        Point clusterAvg = computeClusterAverage(*cluster);
        circle(momentImg, clusterAvg, 3, Scalar(0, 255, 0), CV_FILLED);
    }

    imshow("momentImg", momentImg);
    imwrite("centroids.jpg", momentImg);
    waitKey();

    return 0;
}

vector<Point> getAllTriangleVertices(Mat& img, const vector< vector<Point> >& contours)
{
    vector<Point> approxTriangle;
    vector<Point> allTriangleVertices;
    for(size_t i = 0; i < contours.size(); i++)
    {
        approxPolyDP(contours[i], approxTriangle, arcLength(Mat(contours[i]), true)*0.05, true);
        if(approxTriangle.size() == 3)
        {
            copy(approxTriangle.begin(), approxTriangle.end(), back_inserter(allTriangleVertices));
            drawContours(img, contours, i, Scalar(0, 255, 0), CV_FILLED);

            vector<Point>::iterator vertex;
            for(vertex = approxTriangle.begin(); vertex != approxTriangle.end(); ++vertex)
            {
                circle(img, *vertex, 3, Scalar(0, 0, 255), 1);
            }
        }
    }

    return allTriangleVertices;
}

double euclideanDist(Point a, Point b)
{
    Point c = a - b;
    return cv::sqrt(c.x*c.x + c.y*c.y);
}

vector< vector<Point> > groupPointsWithinRadius(vector<Point>& points, double radius)
{
    vector< vector<Point> > clusters;
    vector<Point>::iterator i;
    for(i = points.begin(); i != points.end();)
    {
        vector<Point> subCluster;
        subCluster.push_back(*i);

        vector<Point>::iterator j;
        for(j = points.begin(); j != points.end(); )
        {
            if(j != i &&  euclideanDist(*i, *j) < radius)
            {
                subCluster.push_back(*j);
                j = points.erase(j);
            }
            else
            {
                ++j;
            }
        }

        if(subCluster.size() > 1)
        {
            clusters.push_back(subCluster);
        }

        i = points.erase(i);
    }

    return clusters;
}

Point computeClusterAverage(const vector<Point>& cluster)
{
    Point2d sum;
    vector<Point>::const_iterator point;
    for(point = cluster.begin(); point != cluster.end(); ++point)
    {
        sum.x += point->x;
        sum.y += point->y;
    }

    sum.x /= (double)cluster.size();
    sum.y /= (double)cluster.size();

    return Point(cvRound(sum.x), cvRound(sum.y));
}

void printPointVector(const vector<Point>& points)
{
    vector<Point>::const_iterator point;
    for(point = points.begin(); point != points.end(); ++point)
    {
        cout << "(" << point->x << ", " << point->y << ")";
        if(point + 1 != points.end())
        {
            cout << ", ";
        }
    }
    cout << endl;
}

I fixed a few bugs in my previous implementation, and cleaned the code up a bit. I also tested it with various resize factors, and it seemed to perform quite well. However, after I reached a quarter scale it started to have trouble properly detecting triangles, so this might not work well for extremely small crosses. Also, it appears there is a bug in the moments function as for some valid clusters it was returning (-NaN, -NaN) locations. So, I believe the accuracy is a good bit improved. It may need a few more tweaks, but overall I think it should be a good starting point for you.

I think my triangle detection would work better if the black border around the triangles was a bit thicker/sharper, and if there were less shadows on the triangles themselves.

Hope that helps!

Answer 2

How about simply determining the autocorrelation as you have a nice periodic pattern in your images.

Let say you have your target image:

And a template image to synchronize to

You can determine the Autocorrelation of both:

In both you can detect the ACF peaks, as in this example

These peaks you can match against each other using the Hungarian algorithm. Here a match is indicated by a white line.

This gives you a set of matching 2D coordinates, which hopefully satisfy the relationship:

x = Ax'

Where A is tranformation matrix with scaling and rotation. Solving it thus gives you the rotation and the scaling between your template and your target image. Translation can then be ascertained via normal cross-correlation with the partly rectified/corrected image and the template image.

Answer 3

I have used a commercial tool called HexSight ( http://www.lmi3d.com/product/hexsight ) in the past in a very similar application. We were pretty happy with its performance and accuracy.

If you're looking for something very rough/basic you can try caluclating the cross correlation between a reference image and the image you're looking at. An alternative (which is what HexSight uses) is to start with some edge detection algorithm before attempting to find matches to the reference image. In either case, you can follow up with some refinement algorithm once you have a rough candidate. Given that you are trying to find some particular target type you may be able to apply some heuristics or take advantage of the specific target with a custom algorithm.

Here's an idea for a custom solution, assuming the crosses are evenly spaced, on a grid, and the image is not too distorted:

• Sum all the rows and columns of the image. This will generate peaks between the rows and columns of the crosses (that are white).
• Determine the image rotation by searching for an angle the maximizes the width and amplitude of those peaks. You can do a binary search between +/- your maximum expected angle.
• Once you have detemined the angle you can now use the centerlines of the peaks to determine the probe locations. You will actually have another narrow/smaller peak right around the center of the targets.

Under the assumptions this should be fairly robust and accurate because it uses information from the entire image. It will not be too sensitive to any local issues, it will even work if a target (or a few) is completely missing or obscured.

Having good optics and illumination are pre-requisities to getting high quality results out of any system. The image you've attches doesn't look that great.