Wolfram Library Archive

Courseware Demos MathSource Technical Notes
All Collections Articles Books Conference Proceedings

Translation and Rotation Invariant 2D Representations for a Neural Network Classifier

Mats Bengtsson
Organization: National Defense Research Establishment
S. Posluk
Journal / Anthology

Försvarets Forskningsanstalt Internal Paper 30692-8.4,3.4
Year: 1993

In this study we present a model for 2D pattern representations which has explicit and simultaneous translation and rotation invariance. in contrast to previous efforts in this direction, we do not rely on center of gravity calculatios and subsequent resampling of the image. Instead, the Fourier-transformed image is used as an input to a rotation invarint operator, i.e., the image is transformed twice. This results in a translation and rotation invarant representation since the rotation operator commutes with the Fourier transform. The invariant features are then used as input to a neural netwrok classifier. The result is a robust and noise insensitive system.

*Applied Mathematics > Computer Science
*Mathematics > Geometry > Plane Geometry