基于改进RT-DETR的航拍图像小目标检测算法  

A Small Target Detection Algorithm in Aerial Images Based on Improved RT-DETR

作　　者：宣岁寒 罗印升[1] 宋伟[1] XUAN Suihan;LUO Yinsheng;SONG Wei(Jiangsu University of Technology,Changzhou 213000,China)

机构地区：[1]江苏理工学院,江苏常州213000

出　　处：《电光与控制》2025年第4期44-51,共8页Electronics Optics & Control

基　　金：江苏省科技计划项目(BY2022134)。

摘　　要：实时、准确地定位与识别航拍图像中飞机、轮船和车辆等目标是进一步决策的根本基础,针对航拍图像中小目标检测存在的效率和精度低等问题,提出了一种基于改进RT-DETR的航拍图像小目标检测算法。首先,通过构建高效的CCFM-P2ASF尺度序列特征融合模块,获得更丰富的语义信息,同时提高对小目标的敏感度;其次,集成灵活性更强的可学习的位置编码,提供更清晰的位置界定;然后,设计更高效的边界框损失函数,减小对目标位置预测的偏差,提供更准确的边界框信息;最后,构建EMA重参数响应模块,从而更有效地提取输入图像特征。实验结果表明:改进后的RT-DETR模型较原始模型参数量减少38.3%,精确率、mAP50和mAP50∶95指标分别提升5.1、5.0和2.2个百分点。对比其他同类主流算法模型,在航拍小目标检测任务中具有更好的检测效果。Real-time and accurate localization and identification of target such as aircraft,ships and vehicles in aerial images is the fundamental basis for further decision-making.Aiming at the issues of low efficiency and accuracy in detecting small targets in aerial images,this paper proposes an improved RT-DETR based small target detection algorithm in aerial images.Firstly,an efficient CCFM-P2ASF scale sequence feature fusion module is constructed to obtain more semantic information and improve the sensitivity to small targets.Secondly,it integrates more flexible learnable positional coding to provide clearer position definition,and then,a more efficient bounding loss function is designed to reduce the deviation of target location prediction and provide more accurate bounding information.Finally,an EMA reparameterization response module is designed to extract input image features more efficiently.The experimental results show that the model size of the improved RT-DETR is reduced by 38.3%,the accuracy,mAP50 and mAP50∶95 is increased by 5.1,5.0,and 2.2 percentage points respectively compared with the original model.The proposed algorithm achieves better detection performance compared with other mainstream algorithm models.

关 键 词：小目标检测 RT-DETR算法 特征融合 定位损失 位置编码 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]