Noisy spectra recognition using a single-layer perception neural network
Michael John Milford · 2002
A variety of single-layer perceptron networks are used to classify minerals based on the Xray spectra generated when they are examined under a scanning electron microscope. The networks tested have different size hidden layers (3, 6, 10 and 15 units) and are trained using different learning rates (0.01, 0.05 and 0.3). The input, hidden and output layers use the linear, hyperbolic and softmax activation functions respectively and the networks are trained using the cross-entropy error-function. The spectral data is pre-processed to yield a feature set that is input into the networks. Two different size feature sets derived from the original spectral data are used. Two levels of artificial noise are added to the original spectra and a further two full size feature sets are derived from these noisy spectra. All four feature sets are then classified by all the networks and the chisquared test . The best performing networks are found to be able to classify the noisest data with 99.6% accuracy. The traditional classification method that uses the chi-squared test is modified to greatly improve its performance, but is found to only classify the same noisy data with 96.2% accuracy. The hidden layer size of the best-performing networks is also found to decrease as the amount of the noise in the data increases.