Wolfram Library Archive

Courseware Demos MathSource Technical Notes
All Collections Articles Books Conference Proceedings

Invariant Representations of 2D Patterns for Neural Networks

Mats Bengtsson
Organization: National Defense Research Establishment
Journal / Anthology


We propose a method for extracting symmetry invariant representations from 2D patterns. The considered symmetry operators are 2D translations and rotations. The extracted information for a specific pattern is invariant under simultaneous 2D translation and rotation. From an initial fourier transform, translation invariant variables are computed. Zernike moments are then used for finding rotation invariant quantities. Contrary to some previous efforts, higher order networks need not be used, implying a much simplified learning procedure. Only a single version of the pattern is needed during the training process, not every translated and rotated copy of it.

*Applied Mathematics > Computer Science