摘要
针对现有目标检测网络很难适应复杂战场环境下的超视距、遮挡、多视角变化等干扰的影响,提出了一种基于多金字塔池化模型的整体嵌套卷积网络,该网络通过引入空洞卷积思想,在保证卷积特征分辨率不变的基础上提高弱小目标的检测精度;同时,本文所提的模型也将通过多孔空间金字塔池化将多尺度上下文特征进行融合,然后在整体嵌套卷积基础上利用装甲目标的形状先验找到感兴趣的目标区域,有效地提升目标的表征能力与抗干扰能力。仿真测试结果表明,本文所提的目标检测识别网络可以有效地提高复杂背景下装甲目标的检测与识别精度。
In view of the fact that the existing object detection network is difficult to adapt to the influence of over-the-horizon,occlusion and multi-view-angle changes in the complex battlefield environment,an improved holistically-nested convolution network based on multi-scale pyramid pooling model is proposed.The network uses context information to improve the detection accuracy of small armored targets.By introducing the idea of dilated/atrous convolution,the detection accuracy of dim-small objects is improved while ensuring the resolution of convolution features remains unchanged.In addition,the proposed model will fuse multi-scale context features through Spatial Pyramid Pooling to enhance the representation ability and anti-jamming performance of the object.Simulation results show that the proposed network can effectively improve the detection and recognition accuracy of armored targets in complex background.
作者
邓磊
李海芳
DENG Lei;LI Hai-fang(Department of Applied Technology,Sichuan Preschool Normal University,Mianyang 621700,China;Institute of Electronic Engineering,China Academy of Engineering Physics,Mianyang 621900,China)
出处
《激光与红外》
CAS
CSCD
北大核心
2022年第2期295-304,共10页
Laser & Infrared
基金
中国工程物理研究院科学技术发展基金项目(No.2019B0203023,No.2017B0403068)
教育部重点职校课题项目(No.DJA170169)资助。
关键词
装甲车辆
目标检测
金字塔池化
上下文特征
视觉语义
弱小目标
armored vehicle
object detection
pyramid pooling
context feature
visual semantics
dim-small objects
