A STATISTICAL APPROACH TO RANK THE MATCHED IMAGE POINTS

Corresponding points matching is a one of the primitive problems in computer vision and image processing which is used in a vast verity of applications such as stereo vision, image registration, object detection, motion analysis and image retrieval. In this paper, we present a new framework to improve the performance of interest points matching using not only the feature descriptors of keypoints, but only the geometrical properties of these points. Recently, although, there have been many attempts to on developing image points matching algorithms, there are some false positives in the results reported. To address the false positive matched pairs, a new approach is presented to detect the outliers based on a matching score between the matched pair points. This score is calculated based on Mahalanobis distance to a learned distribution of matched point’s geometrical properties. Experimental results show that the proposed method is very effective and improves the performance of point matching significantly.