Rotation Invariant Template Matching

In rotation invariant template matching one wants to find finding from the given input image the locations that are similar to the given pattern template, so that the pattern may have any orientation. Template matching is an old but important problem in computer vision and pattern recognition, and thus there have also been many attempts to solve it. Most of the most succesful solutions so far come from the signal processing community, based on fast computation of cross correlation or correlation coefficient. The existing combinatorial approaches have ignored the template rotations. This thesis fills in this gap by presenting the first rotation invariant combinatorial template matching algorithms. The thesis begins by giving the definition, from the combinatorial point of view, of a rotated approximate occurrence of a pattern template in an image. The accuracy of the approximation can be measured by several different distance functions. The consequenses of this definition to the problem complexity are then analyzed. We present several algorithms for solving the problem. There is a trade–off of complexity, efficiency and generality between the algorithms. The simplest one of the algorithms is also the most general one, in terms of distance functions allowed. It evaluates the distance between the pattern and all image positions and pattern rotations in