Abstract
Normalized cross-correlation is the reference approach to carry out templatematching on images. When it is computed in Fourier space, it can handleefficiently template translations but it cannot do so with template rotations.Including rotations requires sampling the whole space of rotations, repeatingthe computation of the correlation each time. This article develops an alternative mathematical theory to handleefficiently, at the same time, rotations and translations. Our proposal has areduced computational complexity because it does not require to repeatedlysample the space of rotations. To do so, we integrate the information relativeto all rotated versions of the template into a unique symmetric tensor template-which is computed only once per template-. Afterward, we demonstrate that thecorrelation between the image to be processed with the independent tensorcomponents of the tensorial template contains enough information to recovertemplate instance positions and rotations. Our proposed method has the potential to speed up conventional templatematching computations by a factor of several magnitude orders for the case of3D images.