Layer-by-layer 3D Reconstruction of Porous Media Based on Cascade Conditional Generative Adversarial Networks

Three-dimensional (3D) images play an important role in the field of digital coring, and the accurate acquisition of 3D images is important for the study of the relevant media structure. Reconstruction of 3D structures from single or multiple two-dimensional (2D) images using computer mathematical modelling methods is currently the main way to obtain 3D structures of research media. The rapid development of deep learning has facilitated the development of 3D reconstruction methods based on generative adversarial networks(GAN). However, this method requires a large number of training samples and can only reconstruct samples with small sizes. In this paper, we propose a cascade conditional generative adversarial network (CCGAN) model based on 2D inter-layer correlation. The model adopts 2D layer-by-layer reconstruction to construct 3D structures, characterises the interlayer variation of 3D structures through joint multi-layer 2D difference images, and learns 3D interlayer correlation using deep learning ideas, which can effectively improve the size and morphological quality of the generated samples. A number of tests have been carried out on isotropic porous media, and the visual inspection and quantitative analysis show that the reconstruction results have good accuracy and diversity, and are in good agreement with the target results.