With mining operations worldwide comprehensively advancing the development of intelligent mines, mine digitalization has become an inevitable trend in the evolution of the mining industry. In this context, efficient and high-precision fracture identification has emerged as a critical prerequisite technology for achieving digital transformation in intelligent mining systems. However, severe environmental noise in deep mines and the limitations of standard pixel-wise loss functions, which ignore the geomechanical continuity of fractures, often lead to fragmented segmentation results. To address this, we introduce the existing SCSegamba architecture as an efficient feature extraction backbone and propose a novel Fracture Orientation Consistency Loss (FOCL) as our core optimization objective. By penalizing directional discrepancies between predicted and actual fracture networks, FOCL explicitly enforces geometric and topological continuity. We evaluated our framework on a custom dataset of 600 high-resolution images from deep underground roadways. The results demonstrate that the SCSegamba-FOCL framework effectively bridges discontinuous fine cracks in low-contrast zones. It achieves a superior Mean Intersection over Union (mIoU) of 85.67% and an F1-Score of 0.869, while maintaining a real-time inference speed of 38 frames per second (FPS) on edge hardware.
Zhang et al. (Sat,) studied this question.