基于Transformer特征金字塔的空中加油锥套实时检测  

Real-Time Detection of Aerial Refueling Drogue Based on Transformer Feature Pyramid

作　　者：李桂林 刘桂华[1] 陈涛 邓豪[1] 唐雪 LI Guilin;LIU Guihua;CHEN Tao;DENG Hao;TANG Xue(School of Information Engineering Southwest University of Science and Technology,Mianyang 621000 China;China Ordnance Equipment Group Automation Research Institute CO.LTD,Mianyang 621000 China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621000 [2]中国兵器装备集团自动化研究所有限公司,四川绵阳621000

出　　处：《电光与控制》2024年第11期75-82,共8页Electronics Optics & Control

基　　金：XX研究共用技术项目。

摘　　要：空中加油锥套的实时检测是实现自主空中加油的重要前提。针对现有目标检测算法在空中加油锥套检测中易受环境干扰导致精度不足的问题,提出了基于Transformer特征金字塔的空中加油锥套实时检测算法。首先,提出一种新的基于池化注意力的Transformer特征金字塔结构TPN对主干特征进行融合,实现更加高效的特征图增强;其次,使用线性注意力进一步降低TPN中注意力机制复杂性,提出轻量化检测模型DNet-LinTPN,减小其80%的内存占用。在自制的空中加油锥套数据集上的实验结果表明,基于TPN的模型在同等条件下精度、速度与模型大小均优于YOLOv7;轻量化检测模型DNet-LinTPN精度达到93.8%较YOLOv7-tiny提升9.4个百分点,参数量减少67.2%,运算量减少45.2%,鲁棒性明显提升。The real-time detection of aerial refueling drogue is an important prerequisite for the realization of autonomous aerial refueling.Since the existing object detection algorithms in aerial refueling drogue detection is susceptible to environmental interferences and may result in insufficient accuracy a real-time detection algorithm for aerial refueling drogue based on the Transformer feature pyramid is proposed.Firstly a new pooling attention-based Transformer feature pyramid structure TPN is proposed for backbone feature fusion to achieve more efficient feature map enhancement.Then linear attention is used to reduce complexity of the attention mechanism in TPN and the lightweight detection model DNet-LinTPN is proposed to reduce the memory consumption by 80%.The experimental results on the self-created air refueling drogue dataset show that the TPN-based model outperforms YOLOv7 in terms of accuracy speed and model size under the same conditions.The lightweight detection model of DNet-LinTPN achieves an accuracy of 93.8%which is a 9.4 percentage point improvement over YOLOv7-tiny with a 67.2%reduction in the amount of parameters and a 45.2%reduction in the amount of operations and the robustness is obviously improved.

关 键 词：自主空中加油 锥套检测 特征金字塔 TRANSFORMER 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]