Extraction And Matching Of The Image Local Invariant Features

Three key points of the image local invariant feature based image matching--theextraction of feature points, the choice of matching strategy and the removing ofmismatches, were analyzed in detail with a complete solution for image matchingpresented. First of all, three detectors including Harris, Harris-Laplace andLaplace-of-Gaussian were studied theoretically and compared in the experiments,which show that Laplace-of-Gaussian is more stable than the other two detectors.Then two detectors were combined as a new detector and more correct matches wereobtained in the experiment. Moreover, compares were conducted on the performanceof different combinations. The result shows that the combination of Harris-Laplaceand Laplace-of-Gaussian gets more correct matches in most instances. Three of themost frequently used matching strategies were compared and the experiment showsthat with the same precision, the matching based on the nearest neighbor distanceratio gets higher recall than the nearest neighbor matching and the threshold basedmatching. In the last step of image matching, random sample consensus (RANSAC)algorithm was applied for the removing of mismatches. Type I error rate and Type IIerror rate got in the experiments are within an acceptable range, which verifies theeffectiveness of the method.Applying the adversarial optimization approach to remove the error matches inimage matching usually causes the removal of correct matches，especially whenmultiple iterations are run．Concerning this drawback，a limitation was put on thenumber of iterations，a subsequent processing was added，and the homography matrixwas re-estimated by using a RANSAC-like method．The experiments show that theimproved method can preserve almost all the correct matches with smaller root meansquare error．And in the aspect of computing speed，the time the improved methodspends is less than half the time the original method spends．...