航空学报 > 2021, Vol. 42 Issue (2): 324223-324223   doi: 10.7527/S1000-6893.2020.24223

基于DNET的空中红外目标抗干扰识别算法

张凯1, 王凯迪2, 杨曦1, 李少毅1, 王晓田1   

  1. 1. 西北工业大学 航天学院, 西安 710072;
    2. 南京南瑞信息通信科技有限公司, 南京 211106
  • 收稿日期:2020-05-15 修回日期:2020-05-29 发布日期:2020-07-06
  • 通讯作者: 杨曦 E-mail:NWPUYX@163.com
  • 基金资助:
    国家自然科学基金(61703337);上海航天科技创新基金(SAST2017-082,SAST2019-081)

Anti-interference recognition algorithm based on DNET for infrared aerial target

ZHANG Kai1, WANG Kaidi2, YANG Xi1, LI Shaoyi1, WANG Xiaotian1   

  1. 1. School of Astronautics, Northwestern Polytechnical University, Xi'an 710072, China;
    2. NARI Information & Communication Technology Company, Nanjing 211106, China
  • Received:2020-05-15 Revised:2020-05-29 Published:2020-07-06
  • Supported by:
    National Natural Science Foundation of China (61703337); Shanghai Aerospace Science and Technology Innovation Fundation (SAST2017-082, SAST2019-081)

摘要: 复杂空战背景下针对人工干扰的博弈是红外空空导弹精确探测制导技术发展面临的瓶颈和核心技术。针对人工干扰对空中红外目标产生的遮蔽、黏连、相似等干扰现象,以及目标机动和相对运动造成的形状、尺度、辐射特性剧烈变化等实际问题,提出一种基于信息特征提取的深度卷积神经网络DNET空中红外图像目标抗干扰识别算法。首先,DNET网络对大尺度特征图像采用密集连接模块,在前部通道保存每一层的网络输出,在网络末端引入特征注意力机制,获得每个特征通道的信息特征识别权重。然后,加入多尺度密集连接模块,并与多尺度特征融合检测结合,提高对大尺度变化情况下的目标特征提取能力。实验结果表明,在伴随红外诱饵干扰的实时检测条件下,红外目标由点目标变化为成像目标,直至充满视场的整个过程中,本文抗干扰识别算法的识别精确度、召回率及识别速度分别达到99.36%、96.95%、132 fps,具备识别精确度和召回率高、识别速度快等优点,并具有良好的鲁棒性。

关键词: 空中红外目标, 抗干扰识别, 特征提取骨干, 卷积神经网络, 密集连接

Abstract: Infrared air-to-air missile anti-interference technology is one of the key technologies to achieve accurate guidance and strike capabilities. Aiming at the practical problems such as shadowing, adhesion, similarity and other interference phenomena caused by artificial interference on aerial infrared targets, and the drastic changes in shape, scale, and radiation characteristics caused by target maneuver and relative motion, this paper proposes an aerial infrared image target anti-interference recognition algorithm based on a feature extraction deep convolutional neural network DNET. Firstly, using dense connections on large-scale feature maps, the DNET network stores the network output of each layer in the front channel. A feature attention mechanism is introduced at the end of the network to obtain the information feature recognition weight of each feature channel. Secondly, a multi-scale dense connection module is added and combined with multi-scale feature fusion detection to improve the ability to extract target features with large-scale changes. Experimental results show that the DNET network can accurately identify the target with the interference of infrared decoy in the process of the infrared target changing from a point target to an imaging target until it fills the field of view. The accuracy, the recall rate, and the recognition speed of DNET reach 99.36%, 96.95%, and 132 fps, respectively, indicating the high recognition accuracy, high recall rate, fast recognition speed, and good robustness of the DNET network.

Key words: aerial infrared targets, anti-interference recognition, feature extraction backbone, convolutional neural networks, dense connection

中图分类号: