Computed tomography (CT) is essential for studying rock microstructures and macroscopic properties, yet its imaging quality is often compromised by complex degradation factors in real-world scenarios. Superresolution (SR) reconstruction techniques for rock CT images aim to enhance image quality significantly and overcome low-resolution limitations. However, most existing SR methods rely on difficult-to-acquire paired training data or employ simulated data that inadequately reflect actual degradation processes, leading to suboptimal performance on real rock CT images. To address the challenge of high-quality three-dimensional (3D) rock CT reconstruction under real degradation conditions, this paper proposes an innovative approach that integrates a physics-driven stochastic degradation model with a lightweight network architecture. Key innovations comprise a physics-driven stochastic degradation model that dynamically utilizes randomized selection and permutation of degradation operations to emulate real-system degradation diversity—enhancing synthetic-to-real distribution alignment; synergistically integrated with a batch normalization-free lightweight 3D network where strategic elimination of batch normalization layers achieves extreme computational efficiency while preserving microstructure-critical spatial fidelity; collectively enabling a robust synthetic-to-real framework trained solely on physics-compliant synthetic data, validated through superior reconstruction quality and physical characteristic preservation. This work provides an efficient and reliable image enhancement solution for high-precision digital core analysis in petroleum exploration and related fields while offering valuable insights for geoscientific image reconstruction.

Close

PrevNext

View figure in article

• Image processing
• Porous media
• Artificial neural networks
• X-ray imaging