contains material from
Template Matching Techniques in Computer Vision: Theory and Practice
Roberto Brunelli © 2009 John Wiley & Sons, Ltd
Linear correspondence measures like correlation and the sum of squared differences between intensity distributions are (technically) non robust. Similarity measures based on the relative ordering of intensity values in image regions have demonstrable robustness both to monotonic image mappings and to the presence of outliers. In spite of the amount of information discarded when only ordinal information is used, the associated similarity measures, such as ordinal correlation of rank distances, preserve good pattern discriminability.
keywords: rank transform, census transform, Spearman correlation coefficient, Kendall correlation coefficient, Bhat-Nayar distance, incremental sign transform.