Hierarchical intelligent lithology recognition for thin section images using enhanced DenseNet

Accurate lithology identification of rock thin sections is a fundamental component of geological exploration and mining research, which provides invaluable geological insights. However, the intricacy and ambiguity of rocks present substantial obstacles to the precise identification of rock thin sections. To tackle this challenge, this paper introduces an enhanced Densely Connected Convolutional Network (DenseNet) model, named “SGCT-DenseNet”, a designation specifically created for this study, designed for thin section rock classification. “SGCT” represents the integration of SoftPool and Gated Channel Transformation (GCT) units, designed to enhance model performance. The GCT unit effectively captures inter-channel relationships, enhancing feature extraction by leveraging channel dependencies. SoftPool, with softmax weighting, replaces conventional pooling methods, reducing information loss while preserving crucial downsampling characteristics. The inclusion of Leaky Rectified Linear Unit (Leaky ReLU) activation mitigates vanishing gradient problems, ensuring stable training. A hierarchical classification approach is employed to categorize the dataset into sedimentary, metamorphic, and igneous rocks for the first-level classification, and further subdivides these into 108 s-level classifications. Data augmentation strategies are employed to enhance the model’s robustness and effectiveness. The test set achieves an accuracy of 99.11% for the primary classification and 98.22% for the secondary classification. Ablation experiments are meticulously designed to validate the effectiveness and rationality of each module within the SGCT-DenseNet model. Furthermore, SGCT-DenseNet outperforms several benchmark models, such as VGGNet16, GoogleNet, ResNet50, MobileNetV3-Large, and VoVNet39, achieving significant accuracy improvements of 11.56%, 5.63%, 6.00%, 3.29%, and 3.12% in the second-level classification task while maintaining a relatively lower number of model parameters. SGCT-DenseNet shows a 3.12% improvement in accuracy compared to the original DenseNet121 model. The results of this study demonstrate that SGCT-DenseNet effectively enhances the model’s characterization ability, improves the accuracy of rock thin section recognition, and offers technical support for the precise classification of rocks, thus holding significant engineering and practical importance.