Hypergraph-Based Network for Accurate Concrete Bridge Surface Damage Instance Segmentation

This study addresses the issues of insufficient semantic information representation and weak higher-order feature relationships in the instance segmentation of concrete bridge surface damages. We propose an improved YOLO 11 model framework, which combines the CBAM attention module and HyperC2Net (Hypergraph-Based Cross-Level and Cross-Position Representation Network) to enhance feature extraction and semantic information modeling, significantly improving the damage detection and segmentation performance. The experiments are conducted on a concrete bridge surface damage dataset containing four common types of damage: cracks, spall, rebar, and corrosion. The performance of the proposed method is compared with existing YOLO models using various evaluation metrics such as mAP50, mAP50-95. The experimental results show that the proposed method improves 2.2% in accuracy over the existing YOLO model.