An Improved Generative Adversarial Network for Micro-Resistivity Imaging Logging Restoration

Technical Summary: An Improved Generative Adversarial Network for Micro-Resistivity Imaging Logging Restoration

Problem Statement
Micro-resistivity imaging logging is a critical technique for analyzing the internal structures of geological formations, such as fractures and lithological boundaries, particularly in complex reservoirs like carbonate and volcanic rocks. However, real-world operational issues—including tool instability, well collapse, and equipment failures—often result in logging images that are noisy, distorted, or contain partially missing regions. Traditional exemplar-based and standard CNN-based restoration methods struggle with the domain-specific high-frequency data and large missing regions inherent in these images. While existing Generative Adversarial Network (GAN) approaches offer promise, they frequently fail to generate realistic textures and maintain semantic coherence when dealing with large areas of missing information in geological contexts.

Methodology
The paper proposes an improved GAN-based framework specifically designed to restore partially missing micro-resistivity imaging logging data. The architecture consists of a generator and a dual-branch discriminator, optimized through a hybrid loss function.

• Generator Network:

  • Base Infrastructure: The generator utilizes a Fully Convolutional Network (FCN) as its foundation.
  • Depthwise Separable Convolutional Residual (DSCR) Blocks: To enhance feature extraction efficiency and reduce computational costs, the model incorporates DSCR blocks. These combine depthwise separable convolutions with residual operations, batch normalization, and Leaky ReLU activations to capture spatial and channel-wise dependencies while preserving semantic and boundary information.
  • Inception Modules: Integrated to expand the network's receptive field and prevent overfitting. These modules employ parallel kernels of varying sizes (1×1, 3×3, and 5×5) with dimensionality reduction to extract multi-scale geological features.
  • Multi-Scale Feature Extraction & Attention: The generator includes a multi-scale feature extraction module with medium and low-resolution branches. It utilizes Feature Extraction Modules (FEM) and Channel Attention Residual (CAR) blocks to highlight significant feature channels and suppress irrelevant ones. Residual connections are used to mitigate information loss during upsampling, and bilinear interpolation is applied to progressively restore spatial resolution.

• Discriminator Network:

  • Dual-Branch Structure: To ensure both global semantic consistency and local texture realism, the framework employs two discriminators: a Global Discriminator that assesses the coherence of the entire reconstructed image, and a Local Discriminator that focuses specifically on the repaired regions and their boundaries. These components engage in adversarial learning to mutually improve the restoration quality.

• Training Strategy and Loss Function:

  • The model is trained using adversarial optimization with a composite objective function comprising five components:
    1. Content Loss ($L_v, L_h$): Calculated for both valid and hole regions using L1 distance and Multi-Scale Structural Similarity (MS-SSIM).
    2. Perceptual Loss ($L_{perc}$): Computed using feature maps from a pre-trained VGG16 network to ensure semantic consistency.
    3. Style Loss ($L_s$): Based on Gram matrix representations to preserve texture and style information.
    4. Adversarial Loss ($L_{adv}$): A weighted combination of global and local discriminator losses to enforce structural coherence at different scales.
  • The total loss is a weighted sum of these components, with empirically determined coefficients.

Key Contributions
The study introduces a novel GAN architecture tailored for the specific challenges of micro-resistivity imaging logging. Its primary contributions include:

1. Depthwise Separable Convolutional Residual Blocks: Designed to increase feature extraction efficiency while lowering computational costs.
2. Inception Modules: Implemented to identify multi-scale geological features and structures effectively.
3. Channel Attention Mechanism: Integrated to highlight significant feature channels within the reconstruction process.
4. Dual Discriminator Framework: A global/local discriminator setup that jointly assesses image coherence and local textures, addressing the limitations of single-discriminator models in handling large missing regions.

Experimental Results
The proposed method was evaluated using real micro-resistivity imaging logging data from the Daqing and Dagang oil fields. Performance was measured using Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR).

• Quantitative Performance: The model achieved an average SSIM of 0.903 across five sets of images with varying missing region sizes. This represents an improvement of approximately 0.3 compared to other similar methods. The highest SSIM scores for training and validation datasets reached 0.964 and 0.972, respectively, with PSNR values peaking at 37.684 dB and 37.436 dB.
• Qualitative Performance: Visual comparisons against GLCIC, WGAN, and PDGAN demonstrated that the proposed method offers superior texture clarity and structural information, particularly in complex geological areas. Unlike WGAN (which produced blurry edges) or PDGAN (which sometimes introduced artifacts near boundaries), the proposed framework maintained better boundary continuity and semantic coherence.

Significance and Claims
The paper claims that the proposed method provides a robust deep learning solution for restoring micro-resistivity imaging log images. By significantly improving semantic structural coherence and preserving texture details, the method facilitates more accurate subsequent geological interpretations, such as fault detection, lithological identification, and formation estimation. The authors assert that this approach contributes to better reservoir characterization and supports decision-making in the oil and gas industry.

Limitations and Future Directions
The authors acknowledge several limitations:

• Generalization: The model was validated primarily on data from two specific regions (Daqing and Dagang), requiring further verification for diverse geological formations.
• Computational Cost: The integration of advanced modules increases computational complexity and memory usage.
• Comparative Scope: The study benchmarked against other GAN configurations but did not compare against Transformer-based architectures or Denoising Diffusion Probabilistic Models (DDPMs).

Future work is proposed to address these gaps by incorporating more diverse geological datasets, exploring hybrid CNN-Transformer architectures for better global dependency capture, and extending the model from 2D image inpainting to 3D volumetric reconstruction.