FB-YOLO (flying bird YOLO): a YOLO-based detector for flying birds in wind farms

The proliferation of clean energy infrastructure has accelerated the global deployment of wind farms, necessitating enhanced mitigation strategies for bird conservation. This study presents a computationally efficient detection algorithm designed to minimize the ecological impact of wind turbines on bird populations. Our primary contributions include three technical innovations: (1) A Dilated Channel Shuffle Squeeze-and-Excitation (DCSSE) attention module that achieves parameter efficiency while preserving detection accuracy; (2) An Adaptive Scale Efficient Intersection over Union (AS-EIoU) loss function optimized for small-target detection; and (3) Integration of PP-LCNet’s lightweight architecture for feature extraction. The proposed FB-YOLO framework demonstrates superior performance compared to the YOLOv8 benchmark across all evaluation metrics. Quantitative analysis reveals significant improvements: precision (+13.57%), recall (+18.86%), mean Average Precision (mAP)50 (+15.48%), and mAP50-95 (+7.89%), accompanied by a model weight reduction of 3.3 MB. Comparative evaluations with alternative detection models confirm FB-YOLO’s dual advantage in achieving higher accuracy with reduced computational complexity. To support this research, we introduce an annotated bird dataset comprising 6,038 high-resolution images. Experimental results on this dataset demonstrate FB-YOLO’s robust performance (95.6% precision, 94.9% recall, 96.4% mAP50, and 43.4% mAP50-95). The optimized architecture demonstrates significant potential for practical implementation in wind farm monitoring systems, offering an effective technological solution for bird collision prevention through real-time detection capabilities.