Abstract: The lithology identification of rock thin sections is an indispensable part of geological analysis, and its precision directly affects the determination of the types, properties, mineral composition, and other microscopic information of subsequent stratigraphic rock, which is of great significance for geological exploration and mineral mining. In order to identify lithology quickly and accurately, an improved MobileNetV2 lightweight model is proposed to address the complex and diverse mineral composition in rock slices, which identifies lithology from a total of 3 700 rock slice images of five types of rocks. The coordinate attention mechanism is embedded in the inverse residual structure of MobileNetV2 to fuse global feature information of multiple minerals in the image. In addition, the classifier in MobileNetV2 is improved to reduce the number of parameters and computational complexity of the model, so as to improve the computing speed and efficiency of the model, and the leaky rectified linear unit (Leaky ReLU) is used as the activation function to avoid the problem of gradient vanishing in network training. Experimental results show that the improved MobileNetV2 model proposed in this paper has a size of only 2.30 MB, and the precision, recall rate, and F1 value on the test set are 91.24%, 90.18%, and 90.70%, respectively, which has high accuracy, and has the best classification effect compared with similar lightweight networks such as SqueezeNet and ShuffleNetV2.

Key words: rock thin section image, lightweight neural network, MobileNetV2, coordinate attention mechanism, lithology identification