RecaT: A CNN-Transformer based network for coal rock image classification

Coal rock image classification is a crucial aspect in geological research. Deep learning based methods have certain advantages in coal rock image classification, but the feature extraction ability of traditional convolutional networks is relatively weak, and increase in accuracy comes at the expense of computational speed. Although the Transformer has a larger model capacity, its generalization ability is not as good as that of convolutional networks. To enhance both accuracy and computational efficiency, thereby better accommodating complex coal rock mining environments, this study introduces a novel hybrid model—RecaT, a neural network architecture that seamlessly integrates CNN and Transformer components. RecaT combines the strengths of CNN and Transformer's respective advantages, so as to learn more comprehensively. To verify the effectiveness of RecaT, this paper compares it with 22 classification methods on the coal rock dataset. Experimental results demonstrate the superiority of RecaT, achieving an impressive accuracy of 91.32% without transfer learning and further boosting to 95.51% with the application of transfer learning on the coal rock dataset. The CNN-Transformer method used in this paper achieves satisfactory results in coal rock image classification without manual feature extraction, which makes the process of coal rock classification more automated and intelligent, improves the classification accuracy and generalization ability, and provides a powerful tool for the field of geological research.