Abstract
Three-dimensional rock pore segmentation is crucial in fields such as geology and petroleum exploration, holding significant importance for oil and gas resource exploration and development. However, existing segmentation methods still present two main limitations: (1) they fail to capture the spatial relationships of pores in 3D when directly applied to 3D rock pore segmentation, inevitably leading to inaccurate segmentation results; (2) they struggle to apply efficiently in resource-constrained scenarios due to the high computational complexity and costly computational demands. To solve the above issues, we propose a novel and lightweight method based on the Mamba architecture, termed LDLK-U-Mamba, for precise and efficient 3D rock pore segmentation. Specifically, we design a Lightweight Dynamic Large Kernel (LDLK) module to capture global contextual information and develop an InceptionDSConv3d module for multi-scale feature fusion and refinement, further yielding more accurate segmentation results. In addition, the Basic Residual Depthwise Separable Block (BasicResDWSBlock) module is proposed to utilize depthwise separable convolutions and the Squeeze-and-Excitation (SE) module to reduce model parameters and computational complexity. Extensive qualitative and quantitative experiments demonstrate that our LDLK-U-Mamba outperforms current mainstream segmentation approaches, validating its effectiveness for rock pore segmentation-particularly in capturing the 3D spatial relationships of pores.