Classification of microscopic images of rock thin sections based on TLCA-ResNet34

Abstract

Identifying microscopic images of rocks is a crucial method for rock identification, playing a vital role in geological exploration and mineral mining. To facilitate the quick classification and identification of rock thin sections under a microscope, a dataset with 3116 microscopic images of 9 types of rock thin sections was developed using publicly accessible network datasets. By adopting the transfer learning method, a context-aware residual block was designed using the coordinate attention(CA) mechanism, and a targeted TLCA-ResNet34 neural network model was developed. This model is capable of extracting deep-layer feature information from entire rock thin section images, thus achieving the classification and identification of microscopic images. The experimental results show that, compared with several other common models, TLCA-ResNet34, while maintaining the light weight of the model, has the best recognition accuracy, recall rate, and Matthews correlation coefficient (MCC) for the microscopy image test set. It can efficiently and accurately identify microscopic images of rocks.