Aligning `dissimilar' Images Directly

This paper introduces a hierarchical algorithm for the registration of corresponding images that do not necessarily have strong global similarity, such as multi-modal images, images with varying illumination (or specular reflection) and images with significant local motion. The method is based on the global maximization of an average local correlation measure, without the generation of similarity surfaces. Fisher's Z-Transformation is used to rectify the correlation coefficient to ensure that additivity between correlation samples is strictly accurate. As a result, the proposed method can handle sizable misalignments in rotations, scale and shear. Direct error functions are also robustified to optimize alignment in the presence of outliers. The result is a completely autonomous system that recovers the global transformation between two images despite substantial visual differences, including contrast reversals, local motion and disjoint image features. The algorithm was successfully tested with a wide variety of images, particularly where conventional frame-to-frame alignment algorithms had failed