Classifying microfossil radiolarians on fractal pre-trained vision transformers

While deep learning techniques, especially image classification using deep learning, continue to evolve, it has been noted that there is a large time gap in applying these techniques in geological studies. Recently, a new architecture called the vision transformer (ViT), which is an alternative to convolutional neural networks (CNN), has attracted considerable attention. In addition, it has been proposed that the pre-training of classification models using mathematically generated images instead of real images, called formula-driven supervised learning (FDSL), achieves a comparative or even higher performance in visual understanding. In this study, we applied these new techniques to the classification of microfossils (radiolarians). Compared with a previous CNN model, the ViT-based model achieved 6-8% higher average precision. On average, the precision of the FDSL pre-trained models was slightly higher than that of the models pre-trained on real images. Therefore, we propose that these techniques may be suitable for image classification in geological tasks.