航空学报 > 2024, Vol. 45 Issue (20): 630702-630702   doi: 10.7527/S1000-6893.2024.30702

场景抽象语义综合模型及其在红外弱小目标检测中的应用

李少毅1, 张雅淇1, 程岳2, 杨曦1(), 张良3, 林健1, 孟中杰1   

  1. 1.西北工业大学 航天学院,西安 710072
    2.中国航空工业集团公司西安航空计算技术研究所,西安 710065
    3.中国空空导弹研究院,洛阳 471009
  • 收稿日期:2024-05-20 修回日期:2024-06-11 接受日期:2024-06-28 出版日期:2024-07-11 发布日期:2024-07-01
  • 通讯作者: 杨曦 E-mail:nwpuyx@163.com
  • 基金资助:
    国家自然科学基金(62273279);国家资助博士后研究人员计划(GZC20232105);西北工业大学博士论文创新基金(CX2024043)

Scene abstract semantic synthesis model and its application in infrared dim and small target detection

Shaoyi LI1, Yaqi ZHANG1, Yue CHENG2, Xi YANG1(), Liang ZHANG3, Jian LIN1, Zhongjie MENG1   

  1. 1.School of Astronautics,Northwestern Polytechnical University,Xi’an  710072,China
    2.Xi’an Aeronautics Computing Technique Research Institute,AVIC,Xi’an  710065,China
    3.China Airborne Missile Academy,Luoyang  471009,China
  • Received:2024-05-20 Revised:2024-06-11 Accepted:2024-06-28 Online:2024-07-11 Published:2024-07-01
  • Contact: Xi YANG E-mail:nwpuyx@163.com
  • Supported by:
    National Natural Science Foundation of China(62273279);Postdoctoral Fellowship Program of CPSF(GZC20232105);Sponsored by Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University(CX2024043)

摘要:

多任务、跨域复杂环境下弱小目标准确、稳定检测在红外预警、搜索与跟踪、精确制导等领域具有重要作用。针对当前算法过于依赖人工设计策略与先验知识,以提取目标信息为目的,对于场景信息的挖掘和利用不充分,性能受限且环境适应性不足等问题,提出了基于场景抽象语义综合的弱小目标检测方法。首先,利用交叉注意力、孪生网络、扩展语义图和自学习双通道的方法设计了4种场景抽象语义综合模型。接着,基于4种语义综合模型设计了基于场景语义综合的弱小目标检测网络,将场景类别语义信息引入到红外弱小目标检测过程实现多类复杂背景下红外弱小目标检测。最后,实验结果表明提出的基于自学习双通道语义综合的弱小目标检测算法精确度到达84.24%,召回率达到89.68%。

关键词: 场景语义, 语义综合, 图像处理, 深度学习, 红外弱小目标检测

Abstract:

Accurate and stable detection of dim and small targets in complex, multi-task, and cross-domain environments plays a critical role in infrared early warning, search and tracking, and precision guidance. Current algorithms overly rely on manually designed strategies and prior knowledge, focusing primarily on target information extraction while insufficiently mining and utilizing scene information. This results in limited performance and inadequate environmental adaptability. To address these issues, this paper proposes a dim and small target detection method based on scene abstract semantic synthesis. First, four scene abstract semantic synthesis models are designed using cross-attention, Siamese networks, extended semantic graphs, and self-learning dual-channel methods. Next, based on these four semantic synthesis models, a dim and small target detection network based on scene semantic synthesis is designed. This network incorporates scene category semantic information into the infrared dim and small target detection process to achieve detection in various complex backgrounds. Finally, experimental results show that the proposed self-learning dual-channel semantic synthesis-based dim and small target detection algorithm achieves an accuracy of 84.24% and a recall rate of 89.68%.

Key words: scene semantics, semantic synthesis, image processing, deep learning, infrared dim and small target detection

中图分类号: