Image recognition of carbonate fossils and abiotic particles based on deep convolutional neural network mode

Abstract Thin sections of carbonate rock offer a more precise and accurate method for identifying mineral characteristics, types of fossils, pore structures, inorganic grain types, and cementation in rocks. Geologists can interpret the depositional environment, diagenesis, and reservoir characteristics of carbonate formations based on the information obtained from thin sections. To accurately identify paleontological fossils in carbonate rocks, geologists need to conduct extensive research on paleontological morphology and undergo extensive training under a microscope for extended periods of time to identify fossils in thin sections. Sometimes, hundreds of carbonate flakes need to be described, which consumes a lot of manpower, resources and money, resulting in limited objectivity and efficiency of the study. Some studies have utilized machine learning to classify carbonate rock particles. However, they have encountered challenges such as using a large number of samples, developing overly complex models, which increases the cost of experiments, and being limited in the recognition of various particle types, particularly rare paleontological types. In this study, we implemented an algorithm based on deep convolutional neural networks to automatically classify paleontological fossils and abiotic particles from thin-section photographs. The model ensures high accuracy in recognition while maintaining a low cost. We trained two classical deep convolutional neural network (DCNN) architectures, VGG-16 and ResNet-18, on the original dataset (1,266 images) and the augmented dataset (6,330 images) containing 11 types, respectively. On the original dataset, the accuracy of the VGG-16 architecture is 79.8%, and the accuracy of the ResNet-18 architecture is 83.9%. On the improved dataset, the VGG-16 architecture achieved 98.8% accuracy, while the ResNet-18 architecture achieved 100% accuracy. This study demonstrates that even small sample datasets can yield strong training results and higher classification accuracies through data augmentation methods. Our findings could provide geologists with an easier and faster way to accomplish the complex and time-consuming task of identifying microscopic flakes.