基于深度学习的砂岩样品三维重构研究

CT scanning and reconstruction provide detailed insights into the internal structure of rocks and enable their quantitative analysis. A critical step in this process involves accurately segmenting different components in the CT images. This study compares and analyzes three deep learning models: the conventional U-Net, the ResUNet enhanced with residual modules, and the ResUNet-TL, which incorporates both residual modules and transfer learning features. Image segmentation using the Weka3D plugin in ImageJ is employed as a baseline for comparison. The analysis reveals that the ResUNet-TL model, leveraging a pre-trained VGG model and deep residual network techniques, outperforms the other models in segmenting complex rock CT images, demonstrating advantages in both accuracy and F1 scores. The ResUNet-TL model is applied to identify fractures in 2D CT images, which are then stacked and reconstructed into a 3D model of the rock sample for quantitative analysis. This approach provides an effective tool for advancing research and applications in rock science.