SEARCHING THE OPTIMAL TRAINING SET FOR NEURAL NETWORK TRAINING

Csaba Stupák - Stanislav Marchevský - Miloš Drutarovský

   Nowadays digital image processing by neural networkS is A very propagated field of signal processing mainly thanks to boisterous development of the VLSI technology. The aim of this work is to find the optimal training set for neural network training. In this case the neural network is used as an optimal nonlinear filter of grayscale images distorted by impulsive noise. In the first part of the paper the issue of image filtering by neural networks is introduced. The second part describes the method of searching the training set. There are used three approaches: searching based on variance, the square error criteria and the statistical cross correlation. In this part the results of searching are listed. The weights found in the searching algorithm are used in the neural filter and the results are discussed.

Keywords: stack filter, neural network, impulsive noise, training set, cross correlation

[full-paper]