DGGAT: Dual-branch gated graph attention transformer for high-accuracy semantic segmentation

Semantic segmentation requires both global context and fine-grained details, yet CNNs struggle with long-range dependencies and Transformers can under-represent low-level structure and be computationally heavy. We propose DGGAT, a dual-branch gated graph attention Transformer: a global branch models long-range context, while a graph-embedded local branch groups semantically related pixels into nodes and applies gated graph attention to sharpen boundaries and small objects. An Attention-Based Feature Selection Fusion Module (ASFM) fuses global and local features to suppress redundancy and balance detail with context. On Cityscapes, DGGAT reaches 84.7% mIoU with a ResNet-101 backbone (82.8% with ResNet-50), and on ADE20K it attains 48.2% mIoU, validating both accuracy and efficiency. These results demonstrate that DGGAT effectively integrates global semantics with fine-detail representation.