Abstract: Abstract: To address the problems of small object scale, complex backgrounds, dense object distribution, and sparse supervision signals in end-to-end detection for Unmanned Aerial Vehicle (UAV) aerial images, an improved small object detection algorithm for UAV aerial images based on RT-DETR, named SOFT-DETR (Small-Object Frequency Transport Detection Transformer), was proposed. First, a Cross-Stage Partial Frequency-Spatial Extraction (CSP-PFSE) module was constructed to enhance high-frequency details, such as object edges, textures, and contours, through joint modeling in the spatial and frequency domains. Then, a Small-Object Dynamic Star Fusion (SDSF) structure was designed. By introducing shallow high-resolution features and combining space-to-depth downsampling, dynamic upsampling, and star feature interaction, the fusion of shallow detailed information and deep semantic information was strengthened. Finally, an Entropic Optimal Transport Regularization (EOTR) auxiliary loss was constructed to alleviate sparse supervision caused by Hungarian one-to-one matching through soft matching. Experimental results on the VisDrone2019 dataset show that the Precision, mean Average Precision mAP@0.5, and mAP@0.5:0.95 of SOFT-DETR reach 63.6%, 49.6%, and 30.7%, respectively, which are 2.9, 2.5, and 1.5 percentage points higher than those of the baseline model. Parameter count and floating-point operations are reduced by 30.2% and 3.7%, respectively. The generalization experimental results show that SOFT-DETR has good adaptability to infrared UAV images and long-distance tiny pedestrian detection scenes.

Key words: Keywords: Unmanned Aerial Vehicle (UAV), aerial photography, small object detection, feature fusion, frequency-domain feature, optimal transport

摘要： 摘 要: 针对无人机（UAV）航拍图像中小目标尺度小、背景复杂、目标分布密集以及端到端检测训练监督信号稀疏等问题，提出一种基于RT-DETR改进的无人机航拍小目标检测算法SOFT-DETR（Small-Object Frequency Transport Detection Transformer）。首先，构建跨阶段部分频率-空间特征提取模块（CSP-PFSE），通过空间域与频率域联合建模增强小目标边缘、纹理和轮廓等高频细节表达；其次，设计小目标动态星型融合结构（SDSF），引入浅层高分辨率特征，并结合空间到深度下采样、动态上采样和星型特征交互机制，增强浅层细节与深层语义融合；最后，构建熵正则化最优传输辅助损失（EOTR），通过软匹配方式缓解匈牙利一对一匹配导致的监督稀疏问题。在VisDrone2019数据集上的实验结果表明，SOFT-DETR的Precision、平均精度均值mAP@0.5和mAP@0.5:0.95分别达到63.6%、49.6%和30.7%，相比基线模型分别提升2.9、2.5和1.5个百分点；参数量和浮点运算量分别下降30.2%和3.7%。泛化实验结果表明，SOFT-DETR在红外无人机图像和远距离微小行人场景中具有较好的适应能力。

关键词: 关键词: 无人机航拍, 小目标检测, 特征融合, 频率域特征, 最优传输