Intelligent coal fracture extraction method using μCT and deep learning

The detailed description of coal fractures is of great significance for the exploration and development of coalbed methane resources. The size, orientation, and distribution density of fractures directly affect the permeability of the coal seam. Accurate identification and extraction of fracture information in coal seams are crucial for revealing the mechanisms of fracture network formation and propagation during hydraulic fracturing. Traditional fracture extraction methods often rely on manual labeling and feature extraction based on certain image processing techniques, which have significant limitations in terms of accuracy and efficiency. This paper proposes a coal fracture extraction method based on Trans-UNet and μCT images. Trans-UNet combines the advantages of Transformer modules and Convolutional Neural Network (CNN), possessing both global feature extraction capability and the ability to capture local details in images, significantly improving segmentation accuracy and network robustness. First, the μCT images of coal samples are pre-processed, including using interpolation methods to improve image quality and data augmentation techniques to increase the number of samples. Subsequently, the processed images are segmented using the Trans-UNet network to extract fracture features, and the segmentation results of different neural network models are compared. The results show that the proposed method outperforms other models on the dataset. The Trans-UNet model achieves an Accuracy of 91.3%, Precision of 89.5%, F1-score of 89.8%, and Intersection over Union (IoU) of 84.0%, significantly outperforming other intelligent models such as U-Net and U-Net++. Combining the fine-grained features of μCT images, the application of the Trans-UNet network to coal fracture extraction tasks is an efficient and accurate solution, providing new insights for image processing tasks in related fields.