Measurement of adulteration in liquids by optical interferograms analysis and deep learning

Abstract

We demonstrate the use of a proposed deep learning model to detect six different degrees of adulteration in alcoholic beverages by classifying interferograms captured through a dual aperture common-path interferometer (DACPI). The proposed two-arm convolutional neural network (TA-CNN) classifier is based on the extraction of linear and non-linear local features by principal components analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), respectively. Then, the features of the reduced vectors are extracted individually with convolutional layers for the classification of three balanced sets of interferograms, with different initial calibration and external perturbation characteristics. In addition, an empirical study of the extracted vectors demonstrates the viability of our interferograms as candidates to be classified by the TA-CNN. The performance of the TA-CNN is compared with modern deep learning models adapted by transfer learning for this specific application. The results show a high average accuracy for all the deep models tested, both for separate and combined sets of 96% and 96.5%, respectively. The proposed TA-CNN is the best performance model, reaching an accuracy of 99.15% for the combined sets. Furthermore, an analysis based on the fast Fourier transform (FFT) corroborates the fact that the relevant information for the classification of interferograms lies in their phase. This approach represents a novel method in optical instrumentation without the use of traditional phase measurement interferometry, the need for highly optimized optical calibration, high-precision optical components, and the obtaining of interferograms datasets with the same DACPI setting up.