Standard Deviation Normalized Summed Squared Difference for Image Registration

Abstract

This paper suggests a new method for image registration, based on a new similarity measure, the standard deviation normalized summed squared difference. Such a similarity measure is explicitly defined on the effective overlap between two images, and the image registration is fulfilled by searching for a global minimum peak of this measure over the entire parameter space. Conceptually the suggested method differs from traditional ones in requiring "zero" image pre- processing like image content dependent or optical system specified spatial windowing, frequency filtering, salient feature detection, or intensity normalization. Without any prior knowledge and image pre-processing, experimental results on a dataset of 90 image pairs with various application backgrounds as well as imaging conditions show a 100% success ratio by our suggestion, which is superior to widely recommended methods of SIFT- stitching (44.4%), linear phase cross-correlation (60%), normalized cross-correlation (54.4%), and some recently top-ranked market products (≈74%).