Improved method of reducing mismatching in manifold template matching

Template matching is highly versatile because its underlying principle is very simple. We speeded up conventional rotary template matching by using manifolds. Manifolds are trajectories that are drawn in multi-dimensional space by using multiple images. However, this method has a problem in that it may incur mismatching by changing the shapes of input image manifolds. Therefore, we reduced mismatching with the previous method by not only focusing on the template manifold's shape but also the manifold's space. We reduced mismatching by creating a feature space that was not excessively affected by shape changes to input image manifolds. The feature space was created so that the template manifold and negative samples were separated in multi-dimensional space. Our new method in experiments using actual images demonstrated a recognition rate of 95.3% and a processing time of 4.7s.