Crack Detection based on the Improved YOLOv11 Model

Crack detection is crucial for the safety assessment of infrastructure. However, due to insufficient multi-scale feature extraction and a lack of contextual information, traditional YOLO algorithms often struggle to detect subtle cracks in complex scenes, resulting in limited detection accuracy. This paper proposes a crack detection method based on YOLO with a multi-scale intelligent agent attention mechanism enhancement. By constructing a proxy node network, this method dynamically integrates semantic information from different hierarchical feature maps. The attention mechanism adaptively suppresses background noise and enhances the feature representation of crack areas. In the feature extraction process, a collaborative strategy with multi-scale receptive fields was introduced to enable the model to capture crack features from local textures to global structures. Experiments using public crack datasets have shown that the proposed module has enhanced precision by $14 \%$ compared to standard attention modules.