ELIMINATION OF TYPE II ERRORS IN FEATURE BASED IMAGE MATCHING: A COMPARATIVE EVALUATION OF GEOMETRIC TRANSFORMATION METHODS

Abstract

Detection and elimination of type II errors has been a fundamental step towards realizing high quality matches in feature-based image matching algorithms. However, choosing the most appropriate outlier detection method that would yield the desired result has been a problem for several users of image matching algorithms. This has made it necessary for studies on the performance of outlier detection methods for elimination of type II errors in feature-based image matching. This study evaluated the performance of three geometric transforms, namely; Affine, Projective, and Conformal transform for detection and elimination of type II errors in feature-based image matching. SURF features were detected and extracted from a stereo pair of images of University of Ilorin senate building acquired at close-range. Initially, 10% of type II errors were matched with 90% correctly identified points using hamming distance. Subsequently, 20, 30, 40, 50, 60, 70, 80, and 90% of type II errors were matched with 80, 70, 60, 50, 40, 30, 20, and 10% of correctly identified points respectively, in stages. At each stage the three transforms were used to detect and eliminate type II errors from the correctly matched points. Performance of the three transforms were evaluated based on their efficacy in detecting and eliminating type II errors, and their robustness in identifying correctly matched points. The results showed that Conformal transform had the best performance with a complete detection and elimination of 90% type II errors from correctly matched points, while Affine and Projective transform eliminated 85% and 75% type II errors respectively. The paper concluded that Projective, Affine, and Conformal transform could be used in feature-based image matching to completely detect and eliminate 75%, 85%, and 90% type II errors respectively from correctly matched points.