Hyperspectral Imaging Model for High-Speed Chemometric Analysis in Industrial Applications

Abstract

Hyperspectral imaging (HSI) has become an effective instrument of the rapid, non-destructive chemometric analysis
in manufacturing. In this paper proposed FejérKorovkin Probabilistic Component Analysis Classification (PCF-PC) is
combined with deep learning to classify materials quickly and correctly. The PCF-PC method incorporates FejerKorokin filtering for the processing of the hyperspectral images. Based on the estimated features, Probabilistic Component analysis is performed for the estimation of the features. With the estimated features, deep learning-based classification is performed for the classification of the industrial materials with the hyperspectral images. The approach is a hybrid between FK smoothing to remove noise, probabilistic extraction of latent features to learn spectral variations and 1D Convolutional Neural networks (CNN) to learn features through a hierarchical approach, and therefore effectively classification of complex hyperspectral data. The framework was tested using the SWIR plastic waste dataset whereby it had overall accuracy of 97.7% and average F1-score of 97.4% and outperformed the traditional PCA, probabilistic PCA, SVM, and standalone deep learner methods. The hybrid model also showed effective processing, and it took less than 5 seconds per hyperspectral cube, which is appropriate to real-time industrial use. The findings indicate the usefulness of probabilistic component analysis with deep learning in high-dimensional hyperspectral data and offers a scalable and accurate solution that is computationally efficient, which can be used in industrial chemometric analyses: plastic sorting, quality control, and environmental monitoring.