Optoelectronic fractal scanning technique for wavelet transform and neural net pattern classifiers

Abstract

A 1D scan which follows Peano's curve to a desired resolution is demonstrated to preserve the 2D proximity relationship and is furthermore shown to be efficient for wavelet transform (WT) processing and artificial neural network pattern recognition. This deterministic fractal sampling method can be implemented in real time using optoelectronic scanning. For example, 2D texture patterns are analyzed by using a 1D WT. Those WT coefficients can be fed into a standard backpropagating neural network for pattern recognition. To speed up training time, a top-down design which generalizes Hopfield's energy landscape approach is given in terms of mini-max pattern classifiers.<>