东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (4): 474-482.DOI: 10.12068/j.issn.1005-3026.2024.04.003

• 信息与控制 • 上一篇    

基于脉冲神经网络的轻量化SAR图像舰船识别算法

谢洪途1, 陈佳兴1, 张琳2, 朱楠楠3   

  1. 1.中山大学·深圳 电子与通信工程学院,广东 深圳 518107
    2.空军预警学院,湖北 武汉 430019
    3.中山大学 系统科学与工程学院,广东 广州 510275
  • 收稿日期:2022-12-09 出版日期:2024-04-15 发布日期:2024-06-26
  • 作者简介:谢洪途(1986-),男,湖南邵阳人,中山大学副教授.
  • 基金资助:
    广东省基础与应用基础研究基金资助项目(2023A1515011588);深圳市科技计划资助项目(202206193000001);国家自然科学基金资助项目(62203465)

Lightweight Ship Recognition Algorithm Based on SNN in SAR Images

Hong-tu XIE1, Jia-xing CHEN1, Lin ZHANG2, Nan-nan ZHU3   

  1. 1.School of Electronics and Communication Engineering,Shenzhen Campus of Sun Yat-sen University,Shenzhen 518107,China
    2.Air Force Early Warning Academy,Wuhan 430019,China
    3.School of Systems Science and Engineering,Sun Yat-sen University,Guangzhou 510275,China. Corresponding author: ZHU Nan-nan,E-mail: zhunn25@mail. sysu. edu. cn
  • Received:2022-12-09 Online:2024-04-15 Published:2024-06-26

摘要:

针对传统方法进行合成孔径雷达(SAR)图像目标识别存在参数多、能耗高等问题,提出了一种基于脉冲神经网络(SNN)的轻量化SAR图像舰船识别算法.首先,利用视觉注意力机制提取SAR图像视觉显著图,采用泊松编码器进行脉冲序列编码,能抑制背景噪声干扰.然后,结合泄漏整合发射(LIF)脉冲神经元和卷积神经网络,构建融合时序信息的SNN模型,能实现SAR图像舰船识别.最后,采用反正切函数作为反向传播时脉冲发射函数的梯度替代函数对SNN模型进行优化,能解决模型难以训练的问题.实验结果表明所提算法具有高精度、少参数、高效率和低能耗等优势,能实现SAR图像高效准确舰船识别.

关键词: 合成孔径雷达图像, 舰船识别, 脉冲神经网络, 轻量化

Abstract:

Due to the more parameters and higher energy-consumption in the traditional methods for the synthetic aperture radar (SAR) image target recognition, this paper proposes a lightweight ship recognition algorithm based on the spiking neural network (SNN) in SAR images. Firsty, the visual attention mechanism is adopted to extract the visual saliency map from SAR images, and the Poisson encoder is adopted for the spike train encode, which can suppress the background noise interference. Then, combined with the leaky integrate-and-fire(LIF) spiking neuron and convolutional neural network, the SNN model integrating the time series information is constructed, which can realize the ship recognition in SAR images. Finally, the SNN model is optimized by using the arctangent function as the surrogate gradient function of the spiking emission function during the backpropagation, which can solve the problem that the SNN model is difficult to train. The experiment results show that the proposed algorithm has higher accuracy, fewer parameters, higher efficiency, and lower energy-consumption, which can achieve efficient and accurate ship recognition in SAR images.

Key words: synthetic aperture radar(SAR) image, ship recognition, spiking neural network (SNN), lightweight

中图分类号: