The Coincidence Similarity Index under Rotation
Résumé
Operations aimed at quantifying the similarity or distance between two given vectors play an especially important role in science and technology, especially in artificial intelligence, pattern recognition, and deep learning. Two families of vector comparison operators can be identified, corresponding to those that are bilinear and those that are non-linear. While the former is mostly founded on the inner product, the latter involves non-linear operations such as maximum, minimum, absolute values, etc. Provided the original data, as well as its pre-processing-aimed at normalization, filtering and feature extraction-it becomes interesting to consider the invariance of the comparison operations to specific types of transformations (e.g. rotation). Though the bilinear comparison operators are often invariant to rotation, the non-linear comparison approaches are almost invariably influenced by rotation. Suggested recently, the coincidence similarity index presents several interesting properties that make it particularly effective for several applications. However, though it is less influenced by rotation when the two compared vectors are similar, this similarity index is not invariant to rotation. In the present work, a procedure is suggested that is capable of making the coincidence similarity index mostly invariant to rotations. The potential of the proposed methodology is illustrates respective to two benchmarking procedures, with encouraging results that complement also the aspect of rotational invariance among the several interesting features of the coincidence similarity index for performing strict comparisons between vectors.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |