Integrating Multi-Scale Feature Sets with Vision Transformer for Enhanced Qualitative Discrimination Pear Browning of Visible-Near Infrared Spectroscopy

Abstract

Near infrared (NIR) spectral analysis is a valuable tool for rapid sample analysis, with the potential for high accuracy in model predictions. However, the choice of spectral variables and their combinations can significantly impact the performance of these models. The integration of multi-scale spectral information through advanced fusion models offers a promising avenue for enhancing NIR analysis capabilities. In this study, we developed a novel qualitative discrimination model for pear browning using visible-near infrared spectra (Vis-NIRS). The model leverages a multi-scale convolutional layer to transform spectra into multi-scale feature sets (MFS), capturing a comprehensive range of variable combinations. By employing a vision transformer (ViT), the model adeptly captures both local and global spectral features. The results demonstrate that the MFS-ViT model demonstrated superior classification performance compared to traditional methods (PLS-DA, RF, 1DCNN) on the tested pear dataset, achieving an accuracy of 99.03% on the validation set. This high level of accuracy was consistently observed across pear datasets of varying sizes. The MFS-ViT model shows potential as a promising method in NIR spectral analysis, offering a new approach for qualitative discrimination of pear browning. Its relatively high classification accuracy and robustness across different dataset sizes suggest that it may have some potential for practical applications in agricultural and food industries. This approach could pave the way for more accurate and efficient quality assessments of perishable goods.