Control point assessment for image registration

Presents several extensions of the basic control point assessment (CPA) algorithm. First, we compare CPA to standard corner detection algorithms and then turn to the question of selecting control points with adequate dispersion, since this is crucial for accurate registration. Two selection methods are proposed. The first consists of clustering the control points via the Lloyd algorithm (S.P. Lloyd, 1957, 1982) followed by selecting the dominant control point in each cluster. This "gold standard" approach produces excellent dispersion but is costly in terms of computational effort. The second selection method consists of subdividing the image and then selecting dominant control points in each subdivision. This is extremely fast and produces results comparable to the Lloyd selection method. The paper concludes with a discussion of how least-squares operator norm information can be coupled with anisotropic diffusion to produce smoothed images without corner degradation.