Super-Resolution Reconstruction of Reservoir Core CT Images via GAN: Advancing Energy Extraction Accuracy

Area of Science:

• Geoscience
• Artificial Intelligence
• Image Processing

Background:

• Accurate reservoir core characterization is vital for oil and gas exploration.
• Super-resolution (SR) techniques offer a cost-effective method to enhance CT image clarity.
• Existing SR algorithms struggle to differentiate high- and low-frequency information in core images.

Purpose of the Study:

• To develop an advanced super-resolution method for reservoir core CT images.
• To improve the characterization of porosity and morphology in geological samples.
• To overcome limitations of pixel-level algorithms in distinguishing image details.

Main Methods:

• Proposed the Multi-scale Fusion Attention Networks (MFAGAN), a Generative Adversarial Network (GAN) based approach.
• Integrated a residual-in-residual fusion attention module for enhanced throat structure discrimination.
• Utilized a multi-scale discriminator and perceptual loss for improved image quality and stability.

Main Results:

• MFAGAN demonstrated superior performance over baseline GANs in LPIPS, SSIM, and subjective image quality.
• Achieved stable and consistent objective and subjective performance on the DeepRock-SR dataset.
• Exhibited strong generalization ability within the dataset distribution, though out-of-distribution evaluation is pending.

Conclusions:

• MFAGAN effectively enhances reservoir core CT image resolution and detail.
• The proposed method shows significant potential for improving geological exploration and development.
• Future work will focus on 3D super-resolution reconstruction for practical applications.