Abstract: The RSD-YOLO algorithm, based on YOLOv8s, has been proposed to address the challenges of low detection performance, severe occlusion, the difficulty of small target feature extraction, and the large number of model parameters inherent in UAV aerial images. Firstly, the Receptive Field Attention (CSP-RFA) module is designed to replace the C2f module to enhance the capability for small target feature extraction, effectively addressing the insensitivity of traditional convolutional operations to positional changes. Secondly, the backbone network and feature fusion network have been made lightweight, a new large-size feature map detection branch has been added, and a Receptive Field Pyramid Network (RFPN) has been proposed to optimize the feature flow direction and improve feature representation. Additionally, the detection head module has been optimized by integrating multi-scale features with a multi-level attention mechanism, and the loss function has been updated to improve the model's detection performance for small targets. In terms of model compression, LAMP pruning is employed to further reduce the number of parameters and the model size. Experimental results demonstrate that the lightweight RSD-YOLO model significantly outperforms the baseline model on the publicly available VisDrone2019 dataset, with a 10.0% increase in precision, a 9.5% increase in mAP0.5 (equivalent to a 24.1% increase), and a 6.9% increase in mAP0.5:0.95 (equivalent to a 29.4% increase). The number of model parameters was reduced from 11.12 million to 4.05 million, representing a 63.6% reduction, and the computational cost was reduced from 42.7 GFLOPs to 25.5 GFLOPs, a 40% reduction. Furthermore, on a newly filtered dataset focusing on small occluded targets, RSD-YOLO showed improvements of 9.1%, 16.1%, and 10.7% on precision, mAP0.5, and mAP0.5:0.95, respectively.

摘要： 针对无人机航拍图像中检测性能低、遮挡严重、小目标特征提取难度大及模型参数量大的问题，提出了基于YOLOv8s的RSD-YOLO算法。首先，设计了感受野注意力模块（CSP-RFA）替代C2f模块，以提升小目标特征提取能力，有效应对传统卷积操作对位置变化不敏感的问题。其次，主干网络和特征融合网络进行了轻量化处理，新增了大尺寸特征图检测分支，并提出了感受野金字塔网络（RFPN），优化特征流动方向，增强特征表达能力。同时，检测头模块经过优化，将多尺度特征集成至具多级注意力机制的检测头中，并替换了损失函数，提升了模型对小目标的检测性能。在模型压缩方面，采用LAMP剪枝，进一步减少了模型的参数量和大小。实验结果表明，轻量化后的RSD-YOLO在公开数据集VisDrone2019上较基线模型有显著提升，精度（P）提高了10.0%，mAP0.5提升9.5%（增幅24.1%），mAP0.5:0.95提高6.9%（增幅29.4%）。模型参数量从11.12M减少至4.05M（减少63.6%），计算量从42.7GFLOPs降至25.5GFLOPs（减少40%）。此外，在仅检测遮挡小目标的新数据集上，RSD-YOLO在P、mAP0.5、mAP0.5:0.95上分别提升了9.1%、16.1%和10.7%。