Image fusion of greyscale images using java and opencv gives strange results

Question

I am currently trying to fuse radiological greyscale images using java and the OpenCV wrapper. I wrote some code to find alike images in a database and fuse them. The fusion part is where I am struggling.This is the method I am struggling with:

public BufferedImage registerImages(BufferedImage source, BufferedImage target) 
throws RegistrationException{

    LinkedList<DMatch> goodMatches = getGoodMatches(source, target);
    if(goodMatches.size() >= 7){
         List<KeyPoint> sourceKeypoints = sourceKeyPointsMat.toList();
        List<KeyPoint> targetKeypoints = targetKeyPointsMat.toList();

        LinkedList<Point> sourcePoints = new LinkedList<>();
        LinkedList<Point> targetPoints = new LinkedList<>();

        for(int i = 0; i < goodMatches.size(); i++) {
            sourcePoints.addLast(sourceKeypoints.get(goodMatches.get(i).queryIdx).pt);
            targetPoints.addLast(targetKeypoints.get(goodMatches.get(i).trainIdx).pt);
        }

        MatOfPoint2f sourceMatOfPoint2f = new MatOfPoint2f();
        sourceMatOfPoint2f.fromList(sourcePoints);
        MatOfPoint2f targetMatOfPoint2f = new MatOfPoint2f();
        targetMatOfPoint2f.fromList(targetPoints);

        Mat homography = Calib3d.findHomography(sourceMatOfPoint2f, targetMatOfPoint2f);            
        Mat transformationResult = new Mat(sourceImageMat.rows(), sourceImageMat.cols(), sourceImageMat.type());                
        Imgproc.warpPerspective(sourceImageMat, transformationResult, homography, transformationResult.size());

        Mat resultImage = new Mat();
        Core.add(transformationResult, targetImageMat, resultImage);

        return mat2BufferedImage(resultImage);
    }
    else{
        throw new RegistrationException();
    }
}

Both the mat2BufferedImage() and the getGoodMatches() have been tested and seem to be valid. There seems to be something wrong with findHomography() and warpPerspective() , since this is the result when I view the transformed (not fusioned) image: https://i.imgur.com/8JRQwtG.png

Does someone have an idea what went wrong? Thank you in advance!

EDIT: So after further investigating, it turns out the transformation-matrix has an extremly strong perspective transformation (values from 400-700). Since the images are pretty identical/have only small differences, I don't understand why this is the result of findHomography() . Are there alternatives to this method?

Answer 1

I'd use drawMatches() to verify there is no outliers in goodMatches . The outliers can completely mess up the computed homography. If you get outliers then you need to use the robust variant of findHomography() .