MarkCapsNet: Road Marking Extraction From Aerial Images Using Self-Attention-Guided Capsule Network

High-definition map building and map navigation systems often require detailed, complete, and up-to-date data of road markings. The real-time and accurate recognition of road markings also serves significantly to the autonomous vehicles. This letter designs a self-attention (SA)-guided high-resolution capsule network to conduct road marking extraction from aerial images. First, by combining the superiorities of capsule formulation and high-resolution network architecture, this model behaves advantageously in providing fine-grained and strong feature semantics for promoting pixel-wise marking extraction accuracy. Furthermore, boosted by the capsule-based SA and adversarial learning mechanisms, the feature encoding quality and robustness are positively enhanced. Quantitative assessments, qualitative inspections, and comparative analyses on two aerial image datasets prove the excellent feasibility and effectiveness of the proposed model in road marking extraction tasks.