航空学报 > 2023, Vol. 44 Issue (22): 629220-629220   doi: 10.7527/S1000-6893.2023.29220

基于多级特征增强融合的红外飞机目标检测方法

张毅, 张焱(), 张宇, 张勇, 刘荻   

  1. 国防科技大学 电子科学学院,长沙  410073
  • 收稿日期:2023-06-26 修回日期:2023-07-17 接受日期:2023-08-30 出版日期:2023-09-07 发布日期:2023-09-06
  • 通讯作者: 张焱 E-mail:atrthreefire@sina.com
  • 基金资助:
    国家自然科学基金(62075239)

Infrared aircraft target detection method based on multi-level feature enhancement fusion

Yi ZHANG, Yan ZHANG(), Yu ZHANG, Yong ZHANG, Di LIU   

  1. College of Electronic Science and Technology,National University of Defense Technology,Changsha  410073,China
  • Received:2023-06-26 Revised:2023-07-17 Accepted:2023-08-30 Online:2023-09-07 Published:2023-09-06
  • Contact: Yan ZHANG E-mail:atrthreefire@sina.com
  • Supported by:
    National Natural Science Foundation of China(62075239)

摘要:

红外飞机目标检测识别在无人侦察、目标制导以及航班监控等领域有广阔的应用前景。为解决红外飞机目标检测识别中目标可用特征少、类间差异小导致的精确识别难度高的问题,提出了一种基于多级特征增强融合的网络模型(MFEFNet)。首先,设计了一种局部与全局特征增强融合模块(LGFE),模块通过设计坐标注意力机制和全局像素注意力机制将深层语义特征和底层细节特征进行增强融合,以自顶向下方式实现深层语义动态指导底层局部特征增强,进而自适应强化网络对红外飞机目标信息的表征;然后,设计了一种全局拓展模块(GEM),在特征金字塔的基础上进一步聚合中间特征图多尺度上下文信息,提高网络对多类别目标的鉴别能力;最后,将GEM模块和LGFE模块进行级联,保持特征长距离依赖关系,产生联合效应,进一步提高网络对目标的表征能力和分类决策能力,有效解决红外飞机目标特征缺失以及类间差异小等问题。此外,本文基于迁移开发了新的地面红外飞机数据集,并通过Gram距离分析了迁移数据与真实红外数据在特征层面上的一致性。在地面红外飞机数据集上的实验结果表明:MFEFNet各个模块是有效的,与其他先进算法相比,本算法对地面红外飞机目标识别的平均均值精度(mAP)提升4.3%以上,识别精确度取得了明显优势。

关键词: 红外飞机目标识别, 机器视觉, 特征增强融合, 全局扩展融合, 注意力机制

Abstract:

Infrared aircraft target detection and recognition has broad prospects for applications in the fields of unmanned reconnaissance, target precision guidance, and flight monitoring. To solve the difficulty in accurate recognition caused by few available features and small difference between classes in infrared aircraft target detection and recognition, a Multi-level Feature Enhanced Fusion Network model (MFEFNet) is proposed. Firstly, a Local and Global Feature Enhancement fusion module (LGFE) is designed. By designing the coordinate attention mechanism and the global pixel attention mechanism, the module enhances and merges the deep semantic features and the low-level detail features. The module implements the top-down mode to dynamically guide the low-level local feature enhancement with deep semantics, and then adaptively strengthens the feature map representation of infrared aircraft target information. Then, a Global Extension Module (GEM) is designed. Based on the feature pyramid, the module further aggregates the multi-scale context information of the intermediate feature map to improve the ability of the network to identify multi-class targets. Finally, the GEM module and LGFE module are cascaded to maintain the long-distance dependence of features and produce a joint effect, which further improves the ability to make feature characterization and target classification decisions and effectively solves the problems of missing features and small differences between classes of infrared aircraft targets. In addition, this paper develops a new ground infrared aircraft dataset based on transfer learning, and analyzes the consistency between transfer data and real infrared data at the feature level by using Gram distance. The experimental results on the ground infrared aircraft dataset show that each module of MFEFNet is effective. Compared with other advanced algorithms, the algorithm proposed can improve the mean Average Precision (mAP) of ground infrared aircraft target recognition by more than 4.3%, and has achieved obvious advantages in recognition accuracy.

Key words: infrared aircraft target recognition, machine vision, feature enhancement fusion, global extension fusion, attention mechanism

中图分类号: