Deep Learning for Materials Recognition from FTIR Images

Abstract

Classifying materials and chemicals from their property and spectra images is important for materials diagnosis and chemical engineering industries in the artificial intelligence (AI) era. In this manuscript, we demonstrate a deep learning process to accurately and fast classify materials and chemicals from their FTIR images. Based on a comprehensive image database consisting of 15816 FTIR images of typical chemicals and materials including dyes, alcohols and polymers, etc., the classification into diverse categories is achieved via a ResNet deep learning neural network coupled with attention module, which achieves high accuracies for the train dataset (precision=0.96, recall=0.97 and F1-score=0.97) and test dataset (precision=0.89, recall=0.89 and F1-score=0.89). The present study serves as a foundation for the materials and chemical recognition from their spectra images and facilitates the machine learning-based automatic image-driven material/chemical recognition and diagnosis process.