基于改进Double-Head RCNN的无人机航拍图像小目标检测算法  

Small target detection algorithm based on improved Double-Head RCNN for UAV aerial images

作　　者：王殿伟 胡里晨 房杰 许志杰[2] WANG Dianwei;HU Lichen;FANG Jie;XU Zhijie(School of Communications and Information Engineering,Xi’an University of Posts&Telecommunications,Xi’an 710121,China;School of Computing and Engineering,University of Huddersfield,Huddersfield HD13DH,UK)

机构地区：[1]西安邮电大学通信与信息工程学院,西安710121 [2]哈德斯菲尔德大学计算机与工程学院,哈德斯菲尔德HD13DH

出　　处：《北京航空航天大学学报》2024年第7期2141-2149,共9页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(62201454);西安邮电大学研究生创新基金(CXJJLY2021058)。

摘　　要：为解决无人机航拍图像中小目标特征信息少且容易被噪声干扰导致现有算法漏检率和误检率高的问题,提出一种改进Double-Head Region-卷积神经网络(RCNN)的无人机航拍图像小目标检测算法。在骨干网络ResNet-50上引入Transformer和可变形卷积(DCN)模块,更有效提取小目标特征信息和语义信息;提出一种基于内容感知特征重组(CARAFE)的特征金字塔网络(FPN)结构模块,解决特征融合过程中小目标被背景噪声干扰而丢失特征信息的问题;在区域建议网络中针对小目标尺度分布特点重新设置Anchor生成尺度,进一步提升小目标检测性能。在VisDrone-DET2021数据集上的实验结果表明:所提算法能提取更具有表征能力的小目标特征信息和语义信息,对比Double-Head RCNN算法,所提算法的参数量增加了9.73×10^(6),FPS损失了0.6,但是AP、AP50和AP75分别提升了2.6%、6.2%和2.1%,APs提升了3.1%。The feature information of small targets in unmanned aerial vehicle aerial images is small and easily interfered with by noise,which leads to the high missed detection and false detection rates of existing algorithms.To address these issues,a small target detection algorithm based on an improved Double-Head region-convolutional neural networks(RCNN)for unmanned aerial vehicle aerial images was proposed.Transformer and deformable convolution networks(DCN) modules were introduced on the backbone network ResNet-50 to extract small target feature information and semantic information more effectively.A feature pyramid network(FPN) structure based on content-aware reassembly of features(CARAFE) was proposed to solve the problem that the small target information is interfered with by the background noise,and the feature information is lost in the process of feature fusion.The generation scale of Anchor was reset according to the characteristics of small target scale distribution in the region proposal network to further improve the small target detection performance.The experimental results on the VisDrone-DET2021 dataset show that the proposed algorithm can extract feature and semantic information of small targets with representational capacity more effectively.Compared with the Double-Head RCNN algorithm,the parameter quantity of the proposed algorithm increases by 9.73×10^(6),and the FPS loss is 0.6.However,AP,AP50,and AP75 increase by 2.6%,6.2%,and 2.1% respectively,and APs increases by 3.1%.

关 键 词：小目标检测 无人机航拍图像 Double-Head RCNN TRANSFORMER 内容感知特征重组 

分 类 号：V279[航空宇航科学与技术—飞行器设计] TP391.41[自动化与计算机技术—计算机应用技术]