Classification of Amino Acids Using Hybrid Terahertz Spectrum and an Efficient Channel Attention Convolutional Neural Network

Abstract

Terahertz (THz) spectroscopy is the de facto method to study the vibration modes and rotational energy levels of molecules and is a widely used molecular sensor for non-destructive inspection. Here, based on the THz spectra of 20 amino acids, a method that extracts high-dimensional features from a hybrid spectrum combined with absorption rate and refractive index is proposed. A convolutional neural network (CNN) calibrated by efficient channel attention (ECA) is designed to learn from the high-dimensional features and make classifications. The proposed method achieves an accuracy of 99.9% and 99.2% on two testing datasets, which are 12.5% and 23% higher than the method solely classifying the absorption spectrum. The proposed method also realizes a processing speed of 3782.46 frames per second (fps), which is the highest among all the methods in comparison. Due to the compact size, high accuracy, and high speed, the proposed method is viable for future applications in THz chemical sensors.