Second‐Order Degradation Modeling and Multiscale Feature Fusion for High‐Fidelity Segmentation of Low‐Quality Digital Rock Images

Digital Rock Physics (DRP) is a critical tool for characterizing rock properties and modeling multiphase flow, but segmenting low‐quality (LQ) rock images remains a key challenge due to partial volume blurring. In this study, we propose a method that leverages a second‐order degradation model to generate physically meaningful synthetic LQ and high‐quality (HQ) image pairs for training a full‐scale connected UNet 3+, enabling accurate segmentation of LQ rock images with varying degradation levels. It captures the overall pore structure in LQ rock images while recovering fine details from HQ Scanning Electron Microscope data. We validate its effectiveness by benchmarking against the watershed‐based segmentation method in terms of porosity, permeability, and pore size distribution. Our method delivers an efficient solution for LQ rock image segmentation, enhancing multiscale pore characterization and petrophysical predictions. This holds significant implications for advancing DRP workflows and deepening the understanding of subsurface rock systems.