SEM-Based Pore Structure Segmentation for Reservoir characterisation Using Semantically Calibrated Anisotropy-Aware Model

Model Architecture

The overall architecture of VLSE-Net (vision-language guided structure-enhanced network) consists of two core modules: LSCM (semantic calibration) and ASRM (anisotropic structural refinement).

LSCM (Language-driven Semantic Calibration Module)

ASRM (Anisotropic Structure Refinement Module)

Overview

Accurate pore segmentation is a fundamental step in digital rock analysis and reservoir microstructural characterisation. In SEM images of rock cores, fluctuating imaging conditions, heterogeneous background textures, and anisotropic/elongated pore morphologies often cause (1) semantic ambiguity between pores and non-pore regions and (2) degraded structural continuity (boundary erosion, fragmentation, and spurious adhesions).

VLSE-Net addresses these challenges via two complementary modules:

• LSCM (Language-driven Semantic Calibration Module): injects text priors and cross-modal constraints to stabilise pore semantics under complex backgrounds.
• ASRM (Anisotropic Structure Refinement Module): applies direction-sensitive structural modelling and adaptive fusion to better recover elongated pores, weak boundaries, and locally connected structures.

Method Summary

1) Language-driven Semantic Calibration Module (LSCM)

LSCM leverages the semantic space of a pre-trained vision-language model (CLIP) as an external knowledge anchor for the concept “pore”.

Core components described in the paper include:

• Text prior injection into visual features (multi-scale semantic modulation along bottleneck/decoder pathways).
• Fine-grained cross-modal interaction (token-level cross-attention at the bottleneck).
• Region-level semantic alignment constraint via an auxiliary alignment loss to encourage cross-modal consistency.

2) Anisotropic Structure Refinement Module (ASRM)

ASRM improves direction-sensitive representation for anisotropic pore structures by introducing multi-branch directional modelling and adaptive gating.

Core ideas described in the paper include:

• A compact representation is processed by parallel directional branches (horizontal/vertical/isotropic) to capture anisotropic patterns.
• A lightweight spatial gating network produces location-adaptive fusion weights.
• A controlled residual refinement stabilises training while enhancing structural continuity.

Results (highlights)

From the abstract in paper/paper-en.tex:

• On the main dataset, VLSE-Net improves Dice by +3.12% over a baseline U-Net.
• Porosity error is reduced by 39.6% (relative reduction).
• The model also shows stability across heterogeneous datasets.

Repository Layout

The open-source training and model code lives under public/:

• VLSENet.py: VLSE-Net model implementation.
• train_VLSENet.py: training script for VLSE-Net.
• train_unet.py: baseline U-Net training script (for comparisons).
• clip/: a lightweight CLIP implementation used by the model.

Quick Start

Environment

Typical dependencies:

• Python 3.10+
• PyTorch + torchvision
• numpy, pillow, tqdm

Dataset format

train_VLSENet.py expects a dataset root with the following subfolders (configurable via CLI flags):

• patch_images/: RGB images
• patch_mask/: binary masks (grayscale)
• text/: optional per-image text prompts (<image_stem>.txt)

Train VLSE-Net

Example:

python train_VLSENet.py \
  --data-root ./dataset

Notes:

• Many model features are enabled by default in the public script (e.g., cross-attention text spatial mode, skip attention, multi-scale fusion, AMP when CUDA is available).
• Learning-rate scheduling is supported via --scheduler plateau (default) which adapts LR based on validation Dice.

Outputs

Training outputs are written under public/reports/ (run-specific folder), including:

• best_*.pt / last_*.pt checkpoints
• training_log.csv / training_log.json
• optional validation visualisations
• optional per-epoch training curves