The accurate registration of image series and exact matching of image parts is essential to numerous problems in computer vision. The framework presented in this paper is a generic matching algorithm suitable for many applications where feature extraction is difficult or inaccurate.
Least squares template matching (LSM) is an area-based matching algorithm. It replaces the conventional multi-stage approach where feature detection is followed by thresholding, binarization and a discrete search. Thus, LSM does not depend on the extraction of binary (also called non-iconic) image features. This is a very important advantage especially in low-contrast and blurred imagery, where feature detection is mostly unreliable. Furthermore, unlike in most correlation methods, the optimum transformation is not searched by testing all possible cases, but approached using an optimization scheme. Assuming that a fair initial guess can be supplied, this is not only faster but also more accurate.