Attention based residual network for medicinal fungi near infrared spectroscopy analysis

Abstract

As an effective technology, near infrared spectroscopy (NIRS) can be widely applied to analysis of active ingredients in medicinal fungi. Multiple regression methods are used to compute the relationship between spectral vectors and ingredient contents. In this paper, an autonomous feature extraction method by using attention based residual network (ABRN) to model original NIRS vectors is introduced. Attention module in ABRN is employed to enhance feature wave bands and to decay noise. Different from traditional NIRS analysis methods, ABRN does not require any preprocessing of artificial feature selections which rely on expert experience. The experiments test ABRN by analyzing original spectrums of medicinal fungi (Antrodia Camphorata and Matsutake), which are from 800 nm to 2500 nm, and predicting active ingredients within them. We compare ABRN with other popular NIRS analysis methods. The root mean square error of Antrodia Camphorata training set (RMSET) and validation set (RMSEV) are 0.0229 g·g-1 and 0.0349 g·g-1 for polysaccharide, and 0.0173 g·g-1 and 0.0189 g·g-1 for triterpene. The RMSET and RMSEV of Matsutake are 0.1343 g·g-1 and 0.2472 g·g-1 for polysaccharide, and 0.0328 g·g-1 and 0.0445 g·g-1 for ergosterol. The R2 (coefficient of determination) of these four ingredients are 0.711, 0.753, 0.847 and 0.807. The results indicate that ABRN has better performance in autonomously extracting feature wave bands from original NIRS vectors, which can decrease the loss of tiny feature peaks.

Keywords

• near infrared spectroscopy
• medicinal fungi
• residual network
• attention mechanism
• deep learning