Measuring image similarity to sub-pixel accuracy

Measuring the similarity between different images is essential for performing registration (alignment) and many related tasks. We show that popular image similarity measures such as the sum of squared differences (SSD), cross correlation (CC), and mutual information (MI), particularly in the presence of noise, may present local optima at half pixel translation intervals when low order interpolators are used to model the discrete images. Several strategies for reducing local optima artifacts during registration are discussed. Results using magnetic resonance image (MRI) data are shown