Investigation of Neural Network Methods for Reconstruction and Classification of Texture Images Under Conditions of Incomplete Information

Technical Summary: Investigation of Neural Network Methods for Reconstruction and Classification of Texture Images Under Conditions of Incomplete Information

Problem Statement
The automated analysis of geological drill-core images is frequently impeded by physical damage, such as plug holes, fractures, and data loss, which create incomplete information scenarios. While deep learning has advanced lithological logging, standard Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) degrade significantly when applied to damaged or "inpainted" data where semantic context is artificially reconstructed. A critical gap exists in understanding the relationship between the perceptual quality of generative image inpainting and the downstream utility of these reconstructed images for semantic classification. Specifically, it remains unclear whether high-fidelity visual reconstruction translates to reliable lithological identification, particularly for minority classes in highly imbalanced datasets.

Methodology
The authors propose an integrated, end-to-end pipeline comprising four sequential stages:

1. Preprocessing (Detection & Segmentation): A Faster R-CNN model localizes core columns, while a Mask R-CNN performs instance segmentation to generate precise binary masks of physical defects (e.g., plug holes).
2. Adversarial Reconstruction (Inpainting): A Contextual Residual Aggregation (CRA) Generative Adversarial Network (GAN) is employed to reconstruct missing regions. The architecture utilizes a dual-path generator: a low-resolution path for global structural context and a high-resolution path for grain-scale texture refinement. An Attention Computing Module (ACM) guides the transfer of high-frequency residuals from valid regions to damaged zones. The training objective combines adversarial loss, pixel-wise L1 reconstruction loss, and an auto-encoding self-reconstruction term to ensure statistical consistency without drifting from valid surrounding textures.
3. Classification: The reconstructed images are cropped into uniform patches and classified using various architectures, including specialized texture networks (Deep Residual Pooling - DRP, Deep Encoding Pooling - DEP), modern CNNs (EfficientNetV2, ConvNeXt), and Vision Transformers (ViT, Swin Tiny).
4. Ensemble Strategy: To address the specific failure modes of individual models, a confidence-based hybrid ensemble is proposed, where a Swin Tiny model overrides EfficientNetV2 predictions for specific minority classes (blank, destructed core, coal, dense rock) when the Transformer's confidence exceeds an 80% threshold.

Experimental Setup and Validation
The study utilizes a dataset of 2,190 high-resolution images (1-meter core sections) from 16 wells in a single deposit in South-Western Kazakhstan, covering nine lithological facies. A critical methodological contribution is the implementation of a strict cross-well validation strategy. Unlike standard random patch splits which suffer from spatial data leakage (yielding artificially high accuracies up to 93.7%), the authors partitioned data at the well level (14 wells for training, 2 distinct wells for testing) to ensure a realistic assessment of generalization.

Key Results

• Reconstruction vs. Classification Gap: The CRA-GAN achieved high perceptual fidelity (PSNR 28.7 dB, SSIM 0.91, FID 74.01). However, this high visual quality did not guarantee classification success. Classification accuracy plateaued at approximately 53% (Mean Class Accuracy ~48%) across most architectures, revealing a "semantic gap" where visually plausible reconstructions contained "hallucinated" features that confused classifiers.
• Architecture Performance:
  • Vision Transformers (ViT): Suffered from severe overfitting (training accuracy ~98% vs. test accuracy 48%) due to a lack of spatial inductive biases in data-constrained geological domains.
  • Swin Tiny: Showed sensitivity to class imbalance, heavily biasing toward dominant clay facies, though it captured large-scale structural cues beneficial for minority classes.
  • EfficientNetV2 & ConvNeXt: Demonstrated the most robust baseline generalization (53% accuracy, 48% MCA), attributed to compound scaling and resistance to overfitting.
• Impact of Inpainting: While overall accuracy remained unchanged (53%) with or without inpainting, CRA-GAN provided targeted improvements for specific minority classes. For instance, recall for "Dense rock" improved 3.5-fold (0.02 to 0.07) and "Medium-grained sandstone" F1-score increased by 0.05, indicating that structural continuity restoration aids feature extraction for complex lithologies.
• Hybrid Ensemble: The confidence-based ensemble (EfficientNetV2 + Swin Tiny) raised the Mean Class Accuracy (MCA) from 48% to 58% by selectively overriding predictions for minority classes. This came at the cost of a slight reduction in overall accuracy (53% to 49%), a trade-off deemed acceptable for improving the detection of economically critical, underrepresented facies.

Significance and Claims
The paper posits that current state-of-the-art generative models, while visually effective, may produce semantically ambiguous features that confound downstream classifiers. The study explicitly quantifies the dependency between reconstruction quality and classification performance, demonstrating that high perceptual metrics are insufficient predictors of utility in automated geological analysis.

The authors position the resulting system not as a fully autonomous classifier, but as a decision-support and screening tool. Given the persistent cross-well accuracy ceiling of 49–53%, the pipeline is designed to automate the labor-intensive first pass over damaged core imagery and prioritize expert attention toward candidate lithofacies, particularly minority classes like coal and dense rock. The work offers a reproducible baseline for non-destructive testing and material science, highlighting the necessity of rigorous cross-well validation to avoid inflated performance claims driven by data leakage.

Limitations and Future Directions
The authors acknowledge limitations, including reliance on RGB imagery (lacking spectral data to resolve transitional lithofacies) and a constrained dataset size from a single deposit. Future work is proposed to integrate hyperspectral data, develop uncertainty-aware classification for synthetic regions, create semantic-aware inpainting metrics, and utilize physics-informed synthetic data generation to mitigate class imbalance.