基于Transformer的航空目标检测算法  

Aviation Target Detection Algorithm Based on Transformer

作　　者：季长清[1] 高志勇 秦静[3] 汪祖民[2] JI Changqing;GAO Zhiyong;QIN Jing;WANG Zumin(College of Physical Science and Technology,Dalian University,Dalian 116622,China;College of Information Engineering,Dalian University,Dalian 116622,China;School of Software Engineering,Dalian University,Dalian 116622,China)

机构地区：[1]大连大学物理科学与技术学院,辽宁大连116622 [2]大连大学信息工程学院,辽宁大连116622 [3]大连大学软件工程学院,辽宁大连116622

出　　处：《无线电工程》2023年第12期2811-2819,共9页Radio Engineering

基　　金：国家自然科学基金青年科学基金项目(62002038)。

摘　　要：近几年,基于深度学习的目标检测算法在航空图像检测任务中得到了广泛应用。针对传统水平目标检测算法无法定位航空图像中大量密集排列的倾斜目标问题,提出了TF-BBAVectors模型算法来实现航空图像中倾斜目标的检测任务。为了避免深度卷积神经网络(Deep Convolutional Neural Network,DCNN)带来的网络退化等问题,使用Transformer结构搭建特征提取网络;针对密集的、小尺度图像目标的问题,采用多尺度特征融合的方法提升检测效果;针对倾斜目标检测的问题,通过边界框边缘感知向量表示任意角度的倾斜目标。在DOTA1.0和SSDD+数据集上的部分测试结果表明,此方法的平均精度分别为72.39%和79.98%,证明了TF-BBAVectors模型算法的有效性。During these years,deep learning-based target detection algorithms have been widely used in aerial image detection tasks.To address the problem that traditional horizontal target detection algorithms cannot locate a large number of densely arranged tilted targets in aerial images,the TF-BBAVectors model algorithm is proposed to implement the task of detecting tilted targets in aerial images.Firstly,to avoid the network degradation and other problems caused by Deep Convolutional Neural Network(DCNN),the Transformer structure is used to build a feature extraction network;secondly,for detection of densely packed,small-scale image targets,a multi-scale feature fusion method is used to improve the detection effect;finally,for the problem of tilted target detection,the tilted targets at arbitrary angles are represented by the bounding box edge-aware vectors.Partial test results on DOTA1.0 and SSDD+datasets show that the average accuracy of this method is 72.39% and 79.98% respectively,which proves the effectiveness of the TF-BBAVectors model algorithm.

关 键 词：深度学习 倾斜目标 航空检测 TRANSFORMER 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]