AirSentinel-YFSNet: Scale-reconstruction enhanced YOLOv8 for UAV intrusion defense

• Proposed AirSentinel-YFSNet, a flight target detection model for UAV intrusion defense based on the YOLOv8n architecture, integrating a lightweight SlimNeck module and a multi-scale adaptive feature reconstruction FRMHead module, achieving a balanced trade-off between detection accuracy and real-time performance. • SlimNeck reduces computational cost by approximately 30% by incorporating VoVNet’s one-shot aggregation and GSConv lightweight convolution, boosting inference FPS from 101 to 133, significantly enhancing system responsiveness. • FRMHead leverages multi-branch receptive field fusion and a distributed focal loss-driven dynamic weighting strategy to improve the number of high-confidence predictions per image (PPI) from 1.7 to over 2.1, and enhances localization accuracy by about 1% to 2%, especially for small UAV targets. • Comprehensive evaluation on the multi-scene AOD-4 dataset—including airplanes, helicopters, drones, and birds—achieved high precision (0.919), notable mean Average Precision at IoU=0.50 (mAP50 = 0.845), and ultrafast inference speed (133 FPS), outperforming baseline YOLOv8n by 2.4% in precision and 33% in speed. • Analyzed limitations arising from lack of extreme weather data (heavy rain, dense fog, nighttime low-light), discussed the trade-offs among precision, recall, and false alarm rates, and outlined future work aiming to enhance environmental adaptability and recall via multimodal fusion and online continual learning. • Conducted sensitivity analyses and multiple repeated random splits to verify stable and reliable model performance, demonstrating optimal parameter settings achieving the best balance of accuracy and speed suitable for embedded or edge computing UAV defense scenarios. • Compared with classical detection algorithms, AirSentinel-YFSNet exhibits superior robustness to challenging backgrounds and extremely small targets, yields higher confidence scores and tighter bounding boxes, and significantly reduces false alarms—providing strong technical support for practical UAV intrusion defense applications. As the threat of drone intrusions continues to increase, this paper proposes a flight target detection model named AirSentinel-YFSNet, aimed to enhance detection accuracy and real-time performance in drone intrusion defense. The model is based on the YOLOv8n (You Only Look Once version 8 nano) network architecture and incorporates two innovative modules: SlimNeck and FRMHead (Feature Reconstruction and Multi-branch Head). SlimNeck employs lightweight top-down multi-scale feature fusion, reducing computational load by approximately 30%, thereby increasing the frame rate (FPS, Frames Per Second) from 101 FPS to 102 FPS. FRMHead introduces distributed focal loss (DFL) and a multi-branch fusion strategy, increasing the number of high-confidence predictions per image (PPI) from 1.7 to 2.1 under high Intersection over Union (IoU) thresholds (such as AP75 and AP95), and improving localization accuracy by approximately 1% to 2%. Combined, AirSentinel-YFSNet achieves comprehensive performance with a precision of 0.919, mean Average Precision at IoU=0.50 (mAP50) of 0.845, mean Average Precision across IoU thresholds from 0.50 to 0.95 (mAP50-95) of 0.554, and a frame rate of 133 FPS on the publicly available AOD-4 (Airborne Object Detection version 4) multi-scene dataset, which covers airplanes, helicopters, drones, and birds. Compared to the baseline YOLOv8n, precision improved by 0.024, and FPS increased by 33%. Although recall slightly decreased (from 0.833 to 0.768), the model demonstrates significant practical value in drone defense scenarios characterized by "low false alarms and high real-time performance." Additionally, this paper focuses on the limitations in the model's generalization ability due to the lack of data under extreme weather conditions (such as heavy rain, dense fog, and nighttime) and discusses the trade-offs between performance metrics, clearly indicating that future work aims to enhance the model's environmental adaptability and recall capability.