Designing a dilated multi-scale attention convolutional neural network to optimize rock micro computed tomography image segmentation: a path to accurate petrophysical and fluid flow characterization

Abstract Semantic segmentation of grayscale rock images is critical in Digital Rock Physics (DRP) for simulating fluid flow in porous media. We propose ARSNet (Accurate Rock Segment Neural Network), a dilated multi-scale pyramid attention convolutional neural network that leverages local and global contextual features to enhance rock micro computed tomography (micro-CT) image segmentation. The architecture integrates multi-scale pyramid, squeeze-and-excitation, and global attention modules, with ResNet50 as an encoder and atrous separable convolution for precise pixel labeling. This improves the estimation of petrophysical and fluid transport properties through robust feature extraction. We compare ARSNet against multi-scale networks such as Pyramid Scene Parsing Network (PSPNet), Pyramid Attention Network (PAN), and DeepLabv3 + using pixel-wise metrics and petrophysical parameters (porosity, Euler number, Lattice Boltzmann-derived permeability). ARSNet achieves superior performance, with a test set F1-score of 99.3% and a blind-sample intersection over union (IoU) of 99.0% (outperforming PSPNet’s 93.4%, PAN’s 98.2%, and DeepLabv3+’s 98.2%). It also provides the closest match for petrophysical and fluid flow properties. These results demonstrate ARSNet’s effectiveness in optimizing segmentation accuracy and parameter estimation. Our findings highlight the potential of multi-scale attention networks for advancing DRP and image processing, offering researchers a powerful tool for accurate rock characterization.