Enhanced feature learning and model lightweighting for real-time UAV detection using FEGP-YOLOv8

In realistic scenarios, unmanned aerial vehicles (UAVs) often present as small targets within complex environmental backgrounds, posing significant challenges for detection. The YOLO series of algorithms have demonstrated proficiency in UAV target detection due to their computational efficiency; however, achieving a balance between accuracy, model size, and detection speed remains challenging. This paper introduces Feature Enhancement and Gradient Propagation-YOLOv8 (FEGP-YOLOv8), a lightweight and high-performance UAV detection model based on YOLOv8. The model optimizes feature representation quality and computational efficiency through a convolutional feature enhancement module, integrating FasterNet and efficient multi-scale attention. The neck network is redesigned to accommodate small target detection tasks, and a Partial Convolution-based Detect head (PDetect) eliminates ineffective computations. Tested on three UAV image datasets, FEGP-YOLOv8 achieves 95.4% mAP50 and 107.9 FPS on the DUT-Anti-UAV dataset with only 3.4 MB model size (reduced by 43% compared with YOLOv8n), higher mAP50, and lower GFLOPs/model size with competitive FPS. This research provides an innovative technical solution for the real-time upgrade of UAV intelligent inspection systems.