HieraASGSegNet: Hierarchical Context Fusion for Semantic Segmentation via Adaptive Superpixel Graph Reasoning

Semantic segmentation remains challenging due to scale variation, occlusion, and ambiguous boundaries in real-world scenes. We present HieraASGSegNet, a hierarchical framework with adaptive superpixel graph reasoning, where each superpixel is treated as a graph node and image-specific edges and weights are learned via multi-scale graph attention. Stacking superpixel graphs across scales enables the model to capture long-range region context while preserving fine boundaries, and a gated fusion module selectively integrates features across scales to balance global semantics with local detail. Extensive experiments show consistent gains over recent lightweight baselines. On Cityscapes, HieraASGSegNet achieves 81.1% mIoU with 57.2 GFLOPs and 34.5 FPS. On ADE20K, it attains 44.3% mIoU with 6.1 GFLOPs and 34.5 FPS, and on COCO-Stuff it reaches 42.2% mIoU under comparable complexity. Compared with strong baselines such as SegFormer, SegNext, VWFormer, and SegMAN, our approach delivers higher accuracy while maintaining a balanced trade-off among model size, computational cost, and inference speed. These results validate adaptive superpixel graph reasoning on hierarchical graphs as an effective and practical recipe for accurate, efficient, and scalable high-resolution semantic segmentation, making HieraASGSegNet well suited to resource-constrained scenarios.