Patch-Based Super-Resolution Generative Adversarial Network for High-Resolution Image Reconstruction in Digital Core Analysis

Abstract Accurate modeling of pore-level flow relies on high-resolution images, typically from X-ray micro-computed tomography (micro-CT). However, such imaging is limited to small samples, often below a viable representative elemental volume. In contrast, core-scale samples provide more representative volumes but at lower resolutions. Bridging this resolution gap is critical for extending laboratory-scale insights to practical field applications. In this study, we present a patch-based super-resolution generative adversarial network (PatchSRGAN) designed to achieve 8× high-fidelity reconstruction of high-resolution sandstone micro-CT images from low-resolution inputs, enabling enhanced pore-scale characterization across scales. Unlike mainstream SRGANs, PatchSRGAN uses a patch-based discriminator to provide more meaningful generator feedback by focusing on local details defined by output patch dimensions during training. We analyzed different model setups defined by the discriminator output dimension against baseline methods, such as SRGAN and bicubic interpolation. Results showed that ensuring high-fidelity, high-resolution image reconstruction (512 × 512 pixels at ~ 4.39 μm/pixel) from low-resolution inputs (64 × 64 pixels at ~ 35.12 μm/pixel) requires a patch-based discriminator output dimension of 8 × 8 or higher for more accurate flow property-related metrics. Model setups were comparatively evaluated using mean absolute error, structural similarity index measure, peak signal-to-noise ratio, recall, and precision. The reconstructed images closely matched real high-resolution images in terms of porosity, preferential flow direction, and pore-size distribution, indicating that our model effectively preserves pore geometry during super-resolution. Moreover, PatchSRGAN demonstrates robustness to noise, maintaining acceptable performance even with degraded low-resolution inputs, which highlights its potential for reconstructing high-resolution images from actual core-scale images that are characteristically noisier.