PotholeBAF-Net: Boundary-Aware Fusion Network with Low-Rank Transformer Bridge for Zero-Shot Cross-Dataset Pothole Segmentation

Cross-dataset pothole segmentation is difficult because illumination, camera characteristics, asphalt texture, and boundary quality vary between training and deployment environments. This study proposes PotholeBAF-Net, an EfficientNet-B3 encoder–decoder network comprising a Low-Rank Transformer Bridge, a Boundary-Aware Fusion (BAF) decoder, and a lightweight Boundary Refinement Head. The bridge uses standard multi-head attention whose pre-softmax per-head affinity matrix is rank-bounded by the query/key dimension; BAF then regulates each encoder skip using a learned soft boundary prior, channel–spatial attention, and spatial–channel gated interpolation with decoder semantics. The model was trained and validated on the Large Public Dataset and evaluated without target-domain training or calibration on the independently collected 188-image Collected Taiz Dataset. Across three independent seeds, PotholeBAF-Net obtained external-test precision of 0.8562±0.1040, recall of 0.4854±0.0716, F1/Dice of 0.6196±0.0395, IoU of 0.4164±0.0365, and MCC of 0.5802±0.0198. Removing the boundary prior, Transformer bridge, or refinement head reduced mean IoU by 0.0763, 0.0685, and 0.0675, respectively. The network contains 12.463 M parameters and requires 4.779 GMACs. These findings support controlled skip transfer as a useful design for target-free pothole segmentation, while the remaining seed variability and recall-dominated failures show that shadows, weak boundaries, and visually ambiguous pavement remain unresolved challenges.