Research on intelligent segmentation method of coal body CT image fracture based on CBAM-UNet

Coal CT image fracture segmentation plays a critical role in fracture information acquisition, and its segmentation accuracy directly determines the quality of pore-fracture spatial reconstruction. Current coal CT image fracture segmentation faces multiple challenges, including complex fracture morphology, difficulties in micro-fracture detection, and similar gray values between fractures and coal matrices, resulting in low efficiency of traditional segmentation methods. Therefore, this paper proposes CBAM-Unet (Convolutional Block Attention Module-Unet), an improved network model for coal body fracture extraction based on U-Net. The CBAM-Unet model leverages the U-Net's symmetric structure and residual connections, enabling complete fracture structure segmentation in complex coal body. The decoder replaces standard convolution with residual modules, enhancing detail and contextual information capture, thereby improving segmentation model performance. The convolutional block attention module is integrated into the model, enhancing fracture feature extraction across channel-spatial dimensions while suppressing coal matrix and mineral interference, effectively capturing cross-dimensional feature correlations to improve segmentation accuracy. The results demonstrate clear advantages in detecting fine and complex-aligned fractures, effectively avoiding mineral and dark coal matrix interference. The CBAM-Unet model achieves accuracy, precision, recall, F1 score, and IoU values of 92.13 %, 95.12 %, 93.67 %, 93.78 % and 92.06 % respectively, outperforming U-Net, DeepLabv3+, and Transformer models, demonstrating superior coal fracture segmentation performance, and providing theoretical support for research on coal mechanical properties, seepage characteristics, pore-fracture spatial reconstruction, and fracture evolution.