three-dimensional reconstruction of porous media images using the vox2vox model in presence of multimineral segmentation information

The evaluation of morphological, topological, statistical, and flow properties of porous media needs high-resolution images of porous media at the pore scale. however, direct access to high-quality tomographic images can be costly and impractical. image reconstruction techniques offer a viable solution for obtaining visually realistic image data. among these methods, deep learning-based approaches have gained significant attention from researchers. however, the utilization of petrographic information from porous media in training such models is rare.in this study, we investigate the impact of incorporating multimineral segmentation information in the training of the vox2vox model. the data used in this model consists of 3d images of a sandstone reservoir along with segmented images into 5 classes, including macro-pores, clay, quartz, feldspar, and high-density minerals. additionally, the vox2vox model has been trained with binary segmented images containing pores and solid phases to compare the effect of multimineral information on model and reconstructed images. incorporating multi-mineral segmentation from the porous media significantly enhances the model’s image reconstruction capabilities, as observed through comparisons of various dynamic and static features. incorporating a five-class dataset has led the model to exhibit lower error at the outset of training, stabilizing after approximately 30 epochs, whereas this point for the model based on the two-class dataset is around 60 epochs. furthermore, the comparison parameters for image quality, specifically ssim score: 0.95, mse score: 0.00013, and psnr score: 38.93, are observed for the first model, while for the second model, they are ssim score: 0.89, mse score: 0.00112, and psnr score: 29.49. the original image has a porosity of 0.229, which increases to 0.233 and 0.244 in the reconstructed images from the models based on five-class and two-class datasets, respectively. additionally, the graphs of other parameters also demonstrate the superiority of the model based on the five-class dataset.

three-dimensional reconstruction of porous media images using the vox2vox model in presence of multimineral segmentation information

the evaluation of morphological, topological, statistical, and flow properties of porous media needs high-resolution images of porous media at the pore scale. however, direct access to high-quality tomographic images can be costly and impractical. image reconstruction techniques offer a viable solution for obtaining visually realistic image data. among these methods, deep learning-based approaches have gained significant attention from researchers. however, the utilization of petrographic information from porous media in training such models is rare.in this study, we investigate the impact of incorporating multimineral segmentation information in the training of the vox2vox model. the data used in this model consists of 3d images of a sandstone reservoir along with segmented images into 5 classes, including macro-pores, clay, quartz, feldspar, and high-density minerals. additionally, the vox2vox model has been trained with binary segmented images containing pores and solid phases to compare the effect of multimineral information on model and reconstructed images. incorporating multi-mineral segmentation from the porous media significantly enhances the model’s image reconstruction capabilities, as observed through comparisons of various dynamic and static features. incorporating a five-class dataset has led the model to exhibit lower error at the outset of training, stabilizing after approximately 30 epochs, whereas this point for the model based on the two-class dataset is around 60 epochs. furthermore, the comparison parameters for image quality, specifically ssim score: 0.95, mse score: 0.00013, and psnr score: 38.93, are observed for the first model, while for the second model, they are ssim score: 0.89, mse score: 0.00112, and psnr score: 29.49. the original image has a porosity of 0.229, which increases to 0.233 and 0.244 in the reconstructed images from the models based on five-class and two-class datasets, respectively. additionally, the graphs of other parameters also demonstrate the superiority of the model based on the five-class dataset.