A Lightweight Multidimensional Feature Network for UAV Small Object Detection

无人机小目标检测在军事、搜救和智慧城市等领域有重要应用，但现有模型参数量大、计算复杂。本研究提出轻量级TriD-UAV检测器，通过构建轻量级特征提取网络TriD-Net和高效特征融合网络DENet，旨在平衡精度与计算效率。TriD-Net利用神经架构搜索引入双分支通用倒残差瓶颈结构提升效率，DENet通过深度通道部分卷积阶段减少计算并融合特征。模型使用解耦检测头和基于Wasserstein距离的损失函数增强小目标检测能力。实验表明，TriD-UAV在VisDrone数据集上实现了良好的性能与轻量化平衡，mAP50~95达到21.3%，参数量和FLOPs显著降低。Drone-based small object detection has significant applications in military operations, search and rescue missions, and smart city initiatives. However, existing models suffer from large parameter sizes and high computational complexity. This study proposes a lightweight TriD-UAV detector, which aims to balance accuracy and computational efficiency by constructing a lightweight feature extraction network (TriD-Net) and an efficient feature fusion network (DENet). TriD-Net employs neural architecture search (NAS) to introduce a dual-branch generalized inverted residual bottleneck structure, enhancing efficiency. DENet reduces computational overhead and fuses features through a depth wise channel-wise partial convolution stage. The model further improves small object detection capability using a decoupled detection head and a loss function based on Wasserstein distance. Experiments demonstrate that TriD-UAV achieves a favorable trade-off between performance and lightweight design on the VisDrone dataset, attaining an mAP50~95 of 21.3% while significantly reducing both parameters and FLOPs.