Accurate segmentation of tunnel lining joints is essential for intelligent tunnel inspection, yet it remains challenging because of texture interference, occlusion-induced discontinuity, and the slender topology of joints in complex in situ environments. This study proposes DASMambaNet, a dedicated segmentation network that integrates Dynamic Adaptive Inhibition Convolution for direction-aware feature extraction and background suppression, a Multi-scale Context-aware Mamba Fusion module for long-range continuity reasoning, and a Lightweight Alignment and Confidence-Modulated Directional Fusion module for boundary-sensitive decoding. Experiments were conducted on the self-constructed TJSD dataset containing 4997 annotated images collected from real tunnel sections. The proposed method achieved 94.12% Dice, 74.63% IoU, 98.86% pixel accuracy, and 88.39% precision, outperforming representative CNN- and Transformer-based methods. Qualitative and ablation results further showed that the network effectively suppresses background noise, restores the continuity of occluded joints, and improves boundary localization. These results indicate that DASMambaNet provides an effective and practical solution for automated tunnel lining joint segmentation and can support subsequent tunnel inspection and structural condition assessment.