A lightweight object detection algorithm for resource-constrained UAVs via multi-module optimization and channel pruning

To tackle the issues of complex structure, high computational complexity, and large parameter size prevalent in most existing object detection algorithms, this study proposes an improved object detection algorithm for UAV images, UAV-YOLO, based on YOLOv8. First, a computationally efficient module C2Faster is designed and integrated into the backbone network to enable streamlined feature extraction. Second, the GSConv convolution module is introduced into the neck network, and the Slim-Neck architecture is adopted to facilitate information exchange between different feature layers while reducing the hardware resource demands. Third, the detection head is redesigned with a shared parameter structure and a convolutional gating unit to further reduce computational complexity and enhance detection accuracy. Finally, a new loss function, F-MPDIoU, is also developed to strengthen the algorithm’s robustness and further improve detection accuracy. To further lightweight UAV-YOLO, the LAMP channel pruning method is employed to eliminate low-impact weight parameters from the algorithm, reducing its complexity and meeting real-time detection requirements. The pruned algorithm, LUAV-YOLO, maintains the n and s variants for diverse deployment scenarios. Experimental results on the DOTAv1.0 dataset demonstrate significant efficiency gains: compared to YOLOv8n, LUAV-YOLOn reduces FLOPs by 56.1%, parameter count by 60%, and improves accuracy by 0.1% while maintaining identical FPS; LUAV-YOLOs achieves even greater reductions, lowering FLOPs by 62.5% and parameters by 70.3% compared to YOLOv8s, with only a 1.3% accuracy drop and stable FPS. Extensive comparisons with state-of-the-art algorithms on DOTAv1.0 and NWPU-VHR10 datasets demonstrate that LUAV-YOLO balances accuracy, real-time speed, and low resource consumption, confirming its applicability to real-world UAV detection tasks.