ViMAF-YOLO: a lightweight detection model integrating multi-scale feature fusion and attention mechanisms

Abstract Wind turbine blades are prone to cracks, peeling, and corrosion during long-term operation due to exposure to harsh environmental conditions, which adversely affect both safety and economic efficiency. To address the limitations of existing detection algorithms in feature representation, feature fusion, and model lightweighting, this study proposes an improved wind turbine blade defect detection model based on the YOLOv11n framework, named ViMAF-YOLO (vision Mamba attention fusion YOLO). Firstly, a cross-layer feature fusion pyramid combined with the CGAFusion module is proposed to effectively mitigate information loss during cross-layer transmission. Secondly, a C3k2-multi-branch re-parameterization convolution feature extraction module is designed to enhance multi-scale feature representation while maintaining computational efficiency. Subsequently, dilated-spatial attention module is introduced to enable accurate localization of defect regions under complex backgrounds and low-contrast conditions. Finally, the efficient vision Mamba module is integrated to achieve model lightweighting, effectively reducing computational costs and improving inference efficiency. Experimental results demonstrate that ViMAF-YOLO achieves an average mAP50 of 88.8% on the wind turbine blade defect dataset, representing a 5.1% improvement over YOLOv11n. Precision and recall increase by 3.9% and 3.3%, respectively, while the FPS reaches 125.4, providing a strong technical foundation for the intelligent maintenance of wind power equipment.