Real-World Super-Resolution for μCT Rock Imaging Using Simple Real-ESRGAN-3D

Department of earth science and engineering, Imperial College London, London, England, United Kingdom

^* Corresponding author: yue.sun25@imperial.ac.uk

Abstract

Computer tomography technology is widely used in geological exploration as a non-destructive 3D imaging modality integrated with computer simulation. However, high-resolution CT scanning is constrained by economic costs and long acquisition times, and higher resolution often reduces the effective field-of-view in practice. Computational super- resolution (SR) offers a scalable alternative by enhancing low-resolution scans and reducing the need for additional acquisitions. The paper presents Simple Real-ESRGAN-3D, a GAN-free, × 2 volumetric SR model for micro-CT (μCT) data. The network is a lightweight residual 3D CNN with trilinear upsampling, trained end-to-end using content-only losses: L1, MSE, and a slice-consistency L1 term that encourages coherence across adjacent slices. This design improves training stability and lowers computational cost while preserving fine pore-scale details. On representative μ CT rock volumes, the method delivers strong SR quality, demonstrating efficient and reliable enhancement from low-resolution inputs. By upgrading image fidelity without adversarial training, Simple Real-ESRGAN-3D enables more efficient pore-scale structural analysis and numerical simulation from standard μCT acquisitions.