Multiscale wavelet-adversarial learning eliminates imaging artifacts in digital rock analysis for reliable reservoir evaluation

Abstract

Generative super-resolution (SR) reconstruction models are widely applied in digital rock research to balance the trade-off between image resolution and the scanning device's field of view. Existing methods often enhance visual details or structural fidelity separately. However, they fail to balance these goals effectively. This failure frequently leads to artifacts that distort porosity and permeability measurements. This paper proposes the Stationary and Discrete Wavelet-Enhanced Generative Adversarial Network (SDWGAN). The model is a hybrid SR approach that integrates two wavelet decomposition methods. This integration addresses the challenge effectively. By integrating multi-scale frequency constraints from wavelet decomposition with adversarial training focused on high-frequency components, our method effectively distinguishes rock boundary details from imaging artifacts. The proposed model adopts a global-local feature integration architecture to preserve fine-grained textures and macroscopic structures. Experimental results on the DeepRock-SR dataset (carbonate, sandstone, coal) demonstrate SDWGAN's enhancements: 0.63–2.12 dB PSNR and 0.01–0.11 SSIM improvements in fidelity, alongside 0.001–0.005 LPIPS and 0.62 NIQE gains in perceptual quality over RGB-domain loss-based models. Simulated seepage results indicate that SDWGAN estimates porosity and permeability with 98 % similarity to the reference images. In conclusion, the proposed model manages the perception-distortion trade-off via frequency domain optimization, ensuring petrophysical consistency between SR results and benchmarks. This approach offers a novel and reliable method for reservoir characterization in the field of petroleum geology.