Herkennen van visuele aspecten van infraroodspectra met neurale netwerken

Abstract

This document describes how neural networks can be trained to classify and recognize infrared spectra. Backpropagation was used as the neural network type. The effect of noise on the recognition capabilities of a network has been investigated by generating 150 spectra with various noise levels out of 3 standard spectra. The trained network appeared to be capable of recognizing spectra correctly up to a noise level of 70%. Recognition appears to be correct up to a noise lebel of 70%. The classifying capabilities of backpropagation of spectra have been studied by training a network with 30 spectra, equally divided over three classes. Fourteen other spectra were used as a control set. Only one spectrum was found to be incorrectly classified. The preliminary conclusion is that neural networks are a useful addition to standard pattern matching techniques, especially for recognizing visual aspects.