In this paper, an image matching algorithm combining a SVD matching approach and scale invariant measure is proposed to relate images with large-scale variations. To obtain a better performance on handling redundant points, we modify the SVD matching approach which enforces the condition of minimal distance between the structures of point patterns at the same time ensures the likeliness of the matched points. Together with the adoption of scale invariant features, the proposed method can match features undergoing significant scale changes and provide a set of matches containing a high percentage of correct matches without any statistical outlier detection.
In this paper, the classical RANSAC approach is considered for
robust matching to remove mismatches (outliers) in a list of
putative correspondences. We will examine the justification for
using the minimal size of sample set in a RANSAC trial and propose
that the size of the sample set should be varied dynamically
depending on the noise and data set. Using larger sample set will
not increase the number of iterations dramatically but it can
provide a more reliable solution. A new adjusting factor is added
into the original RANSAC sampling equation such that the equation
can model the noisy world better. In the proposed method, the noise
variances, percentage of outliers and number of iterations are all
estimated iteratively. Experimental results show that the estimated
parameters are close to the ground truth. The modification can also
be applied to any sampling consensus methods extended from RANSAC.