LASSO Based Similarity Learning of Near-Infrared Spectra for Quality Control

This paper presents a novel method for quality control with similarity learning of near-infrared spectra (NIRS). The product quality is rated by measuring the similarity between the tested products and the standard. To automatic this process, we use mahalanobis distance(MahalD) as the metric to estimate the difference of the segmented NIR spectra that are measured from tobacco product samples and standard references. The features used to score the quality rate are then generated by least absolute shrinkage and selection operator(LASSO), which selects the most related wavelengths from MahalD map. The dimension-reduced MahalD map is then loaded into a more advanced regression learning, like SparseNet, Support Vector Machine(SVM) or Random Forrest(RF) to build a model for quality rate prediction. Results from these regression methods are tested and compared, which show this method is effective and LASSO feature selection can improve the prediction accuracy.