AEL-YOLO: lightweight flying bird detection model in complex backgrounds based on YOLOv8n

Abstract

Bird activity can result in hazards such as bird strikes, power grid short circuits, and reductions in crop yield, thereby posing serious threats to both safety and economic operations. Traditional detection methods have the challenges of insufficient robustness and low computational efficiency. These problems are primary caused by complex background interference, highly deformable bird flight postures, and the limited resources of embedded devices. To address these issues, this study introduces an Adaptive Efficient Lightweight YOLO (AEL-YOLO) based on an optimized YOLOv8n. First, standard convolution in the C2f module is replaced with adaptive kernel convolution. The backbone network with improved module can dynamically adapt to the change in the target scale and morphology of birds during flight. Next, an Efficient Multi Branch and Scale Feature Pyramid Network is designed to enhance feature interaction by integrating lightweight multi-scale convolution with efficient upsampling mechanisms. Finally, a lightweight shared convolutional detection head is developed to substantially reduce computational redundancy. Experimental results show that the AEL-YOLO model achieves a precision of 78.56% and a recall of 63.86%. The proposed model achieves an mAP50 of 68.88%, an mAP75 of 43.11%, and an mAP50-95 of 42.12%. Compared with the baseline model, these results represent improvements of 3.01%, 1.32%, and 1.58%, respectively. In addition, the proposed model reduces the number of parameters to 1.58 million, representing a 47% reduction compared with YOLOv8n and lowers the computational cost to 5.8 GFLOPs. Overall, AEL-YOLO demonstrates robust and efficient performance in complex-background bird detection and is well suited for cost-effective embedded deployment in bird-repellent systems.