Digital rock super-resolution reconstruction with efficient 3D spatial-adaptive feature modulation network

Abstract

High-quality digital rock images are important for studying the micropore structure and flow characteristics of reservoirs, these images should be characterized by high resolution and large field of view (FOV). However, due to the limited imaging capability of the hardware equipment, high resolution and large FOV are often in conflict with each other. The super-resolution (SR) reconstruction techniques, which can extract features from low-resolution images to restore high-resolution details, are currently the main means of improving image resolution. For reconstructing high-quality 3D digital rock images, we propose a new 3D Spatial-Adaptive Feature Modulation Network (3DSAFMN), which inherits the spatial modelling capability of Transformer, fuses the multi-scale input information, and accomplishes the optimization of efficiency and accuracy. The evaluation results show that compared with the current advanced deep learning algorithm, the number of parameters of 3DSAFMN is reduced by 45.5%, the reconstruction speed is increased by 1.70 times, and the reconstruction effect is better. Visualization shows that 3DSAFMN can eliminate noise and blur to the maximum extent and highlight valuable features such as pores, fractures and minerals. Furthermore, we apply 3DSAFMN to external sandstone samples to verify the generalization ability of the model. The pore structure parameters calculation and direct flow simulation demonstrate that the reconstruction results are very close to the real samples in terms of both geometric topology and connectivity. In summary, this work provides an effective and reliable novel model based on deep learning for resolution enhancement of digital rock images.