Analysis of mid-infrared hyperspectral data of black ink by neural network

Abstract

We are studying how to analyze the measurement data of mid-infrared hyperspectral imaging using machine learning. Last year, after preprocessing the data, I attempted to identify ink types from differences in spectral shapes using discriminant analysis, decision trees, and KNN. This year, I tried the same discrimination using a Neural network to see if it increases the rate of correct answers. The spectra obtained by measuring five kinds of black marking pen inks on recycled paper measured by an infrared imaging microscope were used as training data. Using 12 predictors obtained by differentiation of the spectra and principal component analysis, supervised machine learning by neural networks was performed. As a result, the narrow neural network (one hidden all-connecting layer with a size of 10) obtained the highest correct answer rate, and its value slightly increased from 97.6 % to 97.7 %.