<p>Accurate semantic segmentation of grayscale rock micro-CT images is a critical yet challenging step in Digital Rock Physics (DRP) for predicting petrophysical and fluid flow properties. Conventional segmentation methods often introduce user bias and struggle to preserve the pore-space topological integrity, leading to errors in subsequent physical simulations. This study proposes ARSNet (Accurate Rock Segment Neural Network), a novel dilated multi-scale pyramid attention convolutional neural network designed to enhance rock micro-CT image segmentation. The architecture integrates a ResNet50 encoder with atrous convolutions and an empirical decoder incorporating multi-scale pyramid, squeeze-and-excitation, and global attention modules to robustly extract both local and global contextual features for precise pixel-wise labeling. The key novelty of this work lies in its tailored architecture that synergistically combines multiple advanced feature extraction modules specifically to address the challenges of rock image segmentation, moving beyond generic application of existing models. ARSNet’s performance was benchmarked against leading multi-scale networks, including PAN, DeepLabv3+, and PSPNet. ARSNet demonstrated superior performance, achieving a test set F1-score of 99.3% and a blind-sample Intersection over Union (IoU) of 99.0%, outperforming PAN (F1:99.0%, IoU:98.2%), DeepLabv3+ (F1:98.8%, IoU:98.2%), and PSPNet (F1:95.9%, IoU:93.4%). Crucially, beyond pixel-wise metrics, ARSNet provided the closest match to ground-truth petrophysical properties, with the most accurate predictions for porosity, Euler number, and Lattice-Boltzmann-derived absolute permeability. These results demonstrate that ARSNet effectively optimizes both segmentation accuracy and the physical realism of the resulting digital rock models. By delivering both accurate segmentation and reliable fluid flow simulation, the model presents a promising tool for future applications in enhancing reservoir characterization.</p>