Pre-trained U-Net Model for Porosity Segmentation in Carbonate Rock Images (PCP-GAN Project)

This upload contains a pre-trained U-Net segmentation model for automated porosity quantification in thin-section images of carbonate rocks. The model is an essential component of the PCP-GAN (Property-Constrained Pore-scale image reconstruction via conditional Generative Adversarial Networks) framework. Model Purpose: The enhanced U-Net architecture performs semantic segmentation to distinguish pore spaces (blue-dyed epoxy resin) from solid matrix in RGB thin-section images of carbonate formations. This automated porosity calculation serves dual roles: Generating porosity labels during cGAN training data preparation Validating porosity values of synthetically generated images Technical Specifications: Input: 480×480×3 RGB thin-section images Output: Binary segmentation masks (pore vs. solid) Architecture: Enhanced U-Net with attention gates and deep supervision mechanisms Training: Hybrid loss function (Dice coefficient + Binary Cross-Entropy, 0.5:0.5 weighting) Usage: This model is required for running the PCP-GAN framework. Place the file in the saved_models/ directory of the PCP-GAN repository. Related Publication: Sadeghkhani, A., Bennett, B., Babaei, M., & Rabbani, A. (2025). PCP-GAN: Property-Constrained Pore-scale image reconstruction via conditional Generative Adversarial Networks. Related GitHub repository: https://github.com/AliSadeghkhani1990/PCP-GAN