Summary: accepted for publication at EUSIPCO '94
An adaptive invariant transform using neural network techniques
S. Kr¨oner, R. Moratz \Lambda , H. Burkhardt
Technische Informatik I
Technische Universit¨at HamburgHarburg
September 23, 1993
Translation invariant pattern recognition for 1D signals and 2D images can be performed
using the nonlinear fast transforms of the class C T . Unfortunately different backgrounds,
noise or distortions of the object may strongly effect the result of the transform. However
robustness and adaptivity against the mentioned distortions are desired. These properties
are typical for neural nets which suggests to combine them with the class CT .
The RT (rapid transform) --- an element of the class C T --- can be represented by a
static net with simple nodes. We succeeded in realizing this signal flow graph as a neural net
with coupled weights. Now the learning rule backpropagation can be applied for adapting
the RT. First steps with a moving target during the learning process have been successfull
with respect to the separation of different object classes.