Physics Statistical Descriptor-Informed Deep Image Structure and Texture Similarity Metric as a Generative Adversarial Network Optimization Criterion for Three-Dimensional Gray-Scale Core Reconstruction

Digital core refers to the use of three-dimensional physical imaging equipment and mathematical modeling to image the internal microstructure of a core and the use of computers to study the connectivity and pore distribution of the core microstructure. The digital core has emerged as a prominent research avenue in image processing in recent years. A three-dimensional (3D) image can be used to effectively study the microstructure and physical properties of a core. Three-dimensional reconstruction from two-dimensional core images is a crucial advancement in this field. Deep learning is advantageous in image reconstruction. However, when the traditional generative adversarial network (GAN) reconstruction method is adapted for reconstructing gray-scale core images, maintaining texture characteristics is difficult; additionally, it may produce blur artifacts in the GAN-generated gray-scale images. In this study, the physics statistical descriptor-informed deep image structure and texture similarity (PSDI-DISTS) metric, a higher-level metric than the traditional correlation function, is used as the loss function of the network, and a texture feature-constrained GAN (TFCGAN) model is proposed for reconstructing gray-scale core images. In addition, a balanced training strategy integrating L1 and PSDI-DISTS losses is designed to optimize the model performance. The experimental results and seepage simulation analysis showed that TFCGAN can maintain the textural characteristics in the gray-scale core image reconstruction results. Furthermore, the reconstruction results exhibited seepage characteristics similar to those of the target.