Automated classification of natural and synthetic pigments from SEM images using CNN-based deep learning

Abstract

The development of convolutional neural network (CNN) techniques has advanced image classification in various fields, including autonomous vehicles, medical care, and cultural heritage science1, 2–3. This study employed CNN techniques to process large numbers of micrographs, build a dataset of pigment microstructures, and classify their manufacturing processed. As research that automatically classified such micro-morphological features using CNNs remains limited, this study provides an early framework demonstrating the potential of deep learning-based morphological analysis in pigment research within cultural heritage. A dataset of 2654 SEM micrographs was established and analyzed using four architectures—AlexNet, GoogLeNet, ResNet, and VGG. Evaluation metrics such as accuracy, precision, recall, F1-score, and confusion matrix were applied, with VGG16 achieving the highest overall performance. The CNN models reached over 97% accuracy using an 80/20 split for cross-validation. These results indicate that the manufacturing process of a pigment can be inferred rapidly by comparing and classifying its micrographs.