Improved deep residual shrinkage network on near infrared spectroscopy for tobacco qualitative analysis

Abstract

Accurate qualitative analysis of near-infrared (NIR) spectroscopy of tobacco is one of the primary challenges faced by the tobacco industry. The NIR spectrum has the characteristics of high dimensionality, high redundancy, and non-linearity, which leads to the low prediction accuracy of the model. In this paper, we developed an improved deep residual shrinkage network (DRSN) model for enhancing the accuracy of NIR qualitative analysis. The method used gramian angular summation fields (GASF) to convert the NIR spectrum into a two-dimensional image, so that the one-dimensional spectral data was encoded into the geometric structure of the two-dimensional image. The attention module was embedded into the DRSN, which improved the attention of the network to the local bands and peaks of the spectrum. Two adjustment factors were also introduced to achieve adaptive selection of thresholds while reducing the interference of noise in the spectra. In order to investigate the effectiveness and stability of the model, a comparison with support vector machine (SVM), random forest (RF), convolutional neural networks (CNN) and DRSN was provided. The results of the proposed method show improvement in prediction accuracy for both the tobacco grades and producing areas. It is also valuable for the promotion of online analysis and any other applications of qualitative analysis.