Residual networks using multi-task learning algorithm for near-infrared spectroscopy: A case study

Abstract

Near-infrared spectroscopy (NIRS) is a widely used non-destructive detection method known for its efficiency and environmental friendliness. However, the complex and high-dimensional nature of NIRS data presents challenges in accurately correlating spectral information with specific chemical compositions. In this study, an improved ResNet-18 model integrated with multi-task learning to estimate multiple chemical contents from full-dimensional NIRS data is proposed. The present model has been optimized by reducing the number of channels while maintaining the network’s depth to prevent overfitting. The designed model was used to predict four chemical compositions in tobacco, demonstrating superior performance compared with traditional machine learning algorithms. The experimental results indicate that the modified ResNet-18 model offers excellent generalization and predictive accuracy.