<p>This paper proposes FaintNet, a semantic segmentation framework addressing detail degradation, edge ambiguity, and class imbalance in low-illumination scenarios. The framework incorporates a differential convolution-based auxiliary network for high-frequency texture recovery, complemented by a Dual Domain Selection Mechanism (DSM) that adaptively prioritizes critical regions through task-aware feature refinement. To mitigate sample distribution challenges, a hybrid loss that combines cross-entropy, focal loss, and dice loss dynamically balances learning objectives. Based on the LLRGBD dataset, FaintNet achieves 85.51% mIoU and 88.87% mPA, outperforming baseline models by 5.05% and 3.84%, respectively. Experimental results demonstrate enhanced robustness in recovering illumination-degraded features while maintaining segmentation precision, providing an effective solution for low-light semantic segmentation through synergistic integration of explicit edge modeling, adaptive feature selection, and loss equilibrium strategies.</p>