Lithology identification method for rock thin section based on the ghost-mobile-attention fusion model

Abstract The classification of rock thin sections plays an important role in geology, mineral resource exploration, oil and gas evaluation, geological disaster warning, and the advancement of geological science and technology. To achieve efficient and accurate lithologic classification of rock thin sections, an improved lithology identification method is proposed in this study. A model fusion strategy is first adopted to construct a lightweight and efficient feature extraction backbone by taking GhostNet as the base network and integrating the core design concepts of MobileNetV3. Subsequently, an improved ECA (Efficient Channel Attention) mechanism is embedded to enhance feature representation, while dilated convolution is employed to replace conventional convolution in order to expand the receptive field of the model. Finally, the Grams optimizer is introduced to replace the Adam optimizer, improving the optimization process by enhancing generalization performance while effectively controlling memory usage and computational cost. Experimental results demonstrate that the proposed Ghost-Mobile-Attention fusion model achieves hierarchical automatic lithologic identification for three major categories and 108 subcategories of rock thin section images, with classification accuracies of 97.84% and 96.61% on the test set, respectively. Compared with the original model, the improved method reduces the number of parameters by 71.35%, converges faster, and achieves higher lithologic identification accuracy, indicating its effectiveness and practicality for rock thin section classification tasks.