Digital rock reconstruction enhanced by a novel GAN-based 2D-3D image fusion framework

Abstract

Digital rock analysis has become increasingly crucial in earth sciences and geological engineering. However, the multiscale characteristics of rock pores often exceed the capabilities of single-resolution imaging, which is inadequate for a comprehensive description of their characteristics. To address this issue, we introduce a novel multiscale rock image fusion framework based on a generative adversarial network (GAN). This method employs a 3D super-resolution convolutional neural network-based generator and a 2D discriminator to integrate low-resolution 3D images with high-resolution 2D images. Compared to existing methods, our approach directly generates high-resolution 3D data, which offers better continuity. Once trained, the generator can upscale low-resolution inputs to produce corresponding high-resolution outputs, thus completing the feature fusion of images with different resolutions. Experiments were conducted using two distinct datasets, encompassing both pore structure analysis and permeability simulation. The results indicate that the fused and reconstructed digital rocks closely resemble genuine digital rocks in terms of pore structure and flow properties. We have also expanded its application and achieved the fusion of 3D CT images with 2D SEM images. Furthermore, as the impact of low-resolution data decreases with increasing resolution difference. Therefore, it is recommended to select an appropriate scaling factor for effective fusion.