Abstract: To address the issues of identification accuracy and efficiency in rock thin section lithology identification, a rock thin section lithology identification model based on improved GhostNet is proposed. First, an edge guidance module was constructed using multi-directional Sobel operators to effectively extract edge and texture information from rock thin section images. Second, CBAM (Convolutional Block Attention Module) attention mechanism was embedded at key positions within the GhostNet model to enhance the feature extraction capability and further improve the model's identification performance. Finally, the classifier structure was redesigned by replacing the original classifier with a global average pooling layer to reduce the model's parameter count and thereby improve training efficiency. A hierarchical classification approach was adopted to divide the dataset into three major categories and 108 subcategories for coarse-grained and fine-grained identification, respectively. Experimental results demonstrate that the improved GhostNet model achieves a first-level classification accuracy of 98.15% and a second-level classification accuracy of 96.39% on the test set, with a model size of 16.3 MB. Compared with EfficientNetV2, a 1.31 percentage point improvement in first-level classification accuracy was achieved with an 81% reduction in parameters, demonstrating superior classification performance and providing an efficient and accurate solution for lithology identification of rock thin sections.

Key words: rock thin section, rock identification, GhostNet, edge information, Convolutional Block Attention Module(CBAM), attention mechanism, lightweight neural network

摘要： 对于岩石薄片岩性识别中存在的识别精度和效率等问题，提出一种基于改进GhostNet的岩石薄片岩性识别模型。首先，采用多方向Sobel算子构建边缘引导模块，有效提取岩石薄片图像中的边缘和纹理信息。其次，在GhostNet模型的关键位置嵌入CBAM(Convolutional Block Attention Module)注意力机制，强化模型的特征提取能力，进一步提高模型识别效果。最后，重新设计分类器结构，使用全局平均池化层替换原始分类器，减少模型的参数量，进而提高模型训练效率。采用分层分类方法将数据集分为三大类别和108小类别，分别进行粗粒度和细粒度识别。实验结果表明，改进的GhostNet模型在测试集上一级分类准确率为98.15%，二级分类准确率为96.39%，模型大小为16.3 MB。与EfficientNetV2相比，所提模型在一级分类准确率上提高了1.31个百分点，参数量减少了81%，展现出更佳的分类性能，能够为岩石薄片的岩性识别提供高效、准确的解决方案。

关键词: 岩石薄片, 岩性识别, GhostNet, 边缘信息, CBAM注意力机制, 轻量级神经网络