<p>High-resolution digital rock images provide the foundation for pore-scale reservoir characterization, seepage simulation, and multi-scale flow analysis. However, the inherent trade-off between field of view and spatial resolution in micro-computed tomography limits the acquisition of high-resolution images over macro-scale rock samples. Deep learning-based super-resolution offers a way to mitigate this limitation by learning mappings from low-resolution to high-resolution images, aiming to recover fine-scale pore–throat structures while preserving a large field of view. Yet existing super-resolution methods still struggle with the strong structural heterogeneity and frequency characteristics of rock images, often producing blurred pore boundaries or over-smoothed mineral textures. This paper proposes a high/low-frequency guided structure-aware super-resolution network (HLF-SASR), which combines a component-aware backbone with a stacked hourglass backbone with an atrous spatial pyramid pooling (ASPP)-based multi-scale context module, a high/low-frequency collaborative attention module, and a temperature-scaled multi-expert fusion strategy. Experiments on a public digital rock dataset using two-dimensional sandstone and carbonate slices, where <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$200\times 200$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>200</mn> <mo>×</mo> <mn>200</mn> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$400\times 400$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>400</mn> <mo>×</mo> <mn>400</mn> </mrow> </math></EquationSource> </InlineEquation> image pairs are used as low-resolution and high-resolution inputs for <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$2\times $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>2</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation> super-resolution, show that HLF-SASR consistently surpasses representative super-resolution models such as enhanced deep super-resolution network (EDSR), residual channel attention network (RCAN), and multi-attention super-resolution neural network (MASR) in terms of peak signal-to-noise ratio and structural similarity index. Visual comparisons further demonstrate that the proposed network produces clearer and more continuous pore–throat boundaries, better preserves narrow fractures and sharp-corner structures, and avoids over-sharpening mineral matrices or introducing pseudo-textures, thereby providing more reliable reconstructions for digital rock analysis.</p>

HLF-SASR: A High–Low Frequency Guided Structure-Aware Super-Resolution Network for Digital Rock Images

期刊论文
发表日期：

2026年4月2日

查看全文

Tianyu Gao,
Chenyang Zhu,
Zhenwei Niu,
Mingjie Li,
Yunxin Xie,
Fang Wang

摘要

High-resolution digital rock images provide the foundation for pore-scale reservoir characterization, seepage simulation, and multi-scale flow analysis. However, the inherent trade-off between field of view and spatial resolution in micro-computed tomography limits the acquisition of high-resolution images over macro-scale rock samples. Deep learning-based super-resolution offers a way to mitigate this limitation by learning mappings from low-resolution to high-resolution images, aiming to recover fine-scale pore–throat structures while preserving a large field of view. Yet existing super-resolution methods still struggle with the strong structural heterogeneity and frequency characteristics of rock images, often producing blurred pore boundaries or over-smoothed mineral textures. This paper proposes a high/low-frequency guided structure-aware super-resolution network (HLF-SASR), which combines a component-aware backbone with a stacked hourglass backbone with an atrous spatial pyramid pooling (ASPP)-based multi-scale context module, a high/low-frequency collaborative attention module, and a temperature-scaled multi-expert fusion strategy. Experiments on a public digital rock dataset using two-dimensional sandstone and carbonate slices, where $200\times 200$ $200 \times 200$ and $400\times 400$ $400 \times 400$ image pairs are used as low-resolution and high-resolution inputs for $2\times $ $2 \times$ super-resolution, show that HLF-SASR consistently surpasses representative super-resolution models such as enhanced deep super-resolution network (EDSR), residual channel attention network (RCAN), and multi-attention super-resolution neural network (MASR) in terms of peak signal-to-noise ratio and structural similarity index. Visual comparisons further demonstrate that the proposed network produces clearer and more continuous pore–throat boundaries, better preserves narrow fractures and sharp-corner structures, and avoids over-sharpening mineral matrices or introducing pseudo-textures, thereby providing more reliable reconstructions for digital rock analysis.