Lu Yong, Sun Minghao
Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE, School of Information Engineering, Minzu University of China, Beijing, 100081, China.
Sci Rep. 2025 Jan 8;15(1):1340. doi: 10.1038/s41598-025-85488-z.
Detecting small targets in UAV remote sensing images is challenging for traditional lightweight methods due to difficulty in feature extraction and high background interference. We propose LPS-YOLO, which improves small target feature extraction while reducing computational complexity by replacing the Conv backbone with SPDConv to retain fine-grained features. LPS-YOLO introduces the SKAPP module for better feature fusion and incorporates the E-BiFPN and OFTP structures to efficiently preserve and transfer backbone information. Evaluation of the VisDrone2019 dataset shows a 17.3% increase in mean Average Precision (mAP) and a 42.5% reduction in parameters compared to the baseline. Additional experiments on the DOTAv2 dataset demonstrate the model's robustness, with a 14.5% improvement in F1 score and a 14.9% increase in mAP over YOLOv8-n. LPS-YOLO offers an effective solution for multi-target detection in UAVs.
对于传统的轻量级方法而言，在无人机遥感图像中检测小目标具有挑战性，这是因为特征提取困难且背景干扰大。我们提出了LPS-YOLO，它通过用SPDConv替换卷积主干来改进小目标特征提取，同时降低计算复杂度，以保留细粒度特征。LPS-YOLO引入了SKAPP模块以实现更好的特征融合，并结合了E-BiFPN和OFT P结构来有效地保存和传递主干信息。对VisDrone2019数据集的评估表明，与基线相比，平均精度均值（mAP）提高了17.3%，参数减少了42.5%。在DOTAv2数据集上的额外实验证明了该模型的鲁棒性，与YOLOv8-n相比，F1分数提高了14.5%，mAP提高了14.9%。LPS-YOLO为无人机中的多目标检测提供了一种有效的解决方案。
