Journal of Applied Optics, Volume. 43, Issue 5, 893(2022)

Hyperspectral target tracking based on attention mechanism and additive angular margin loss

Zonghan SHI and Haitao ZHAO*
Author Affiliations
  • [in Chinese]
  • show less
    Figures & Tables(15)
    Structure diagram of network model
    Structure diagram of attention mechanism
    Image entropy of each band of original hyperspectral images
    Optimum index factor curves of band selection
    Display of each band of hyperspectral image samples
    Spectral curves of target features in hyperspectral images
    Flow chart of online tracking algorithm
    Visualization of attention mechanism
    Comparison of feature maps and attention maps of target and backgrounds
    Tracking results of proposed algorithm
    Comparison of results with different tracking algorithms
    Comparison curves of accurate rate and success rate of different networks
    • Table 1. Loss calculation of AAML

      View table
      View in Article

      Table 1. Loss calculation of AAML

      算法1AAML损失计算
      输入:高光谱图像数据集正负样本及其标签 {(xi,yi),i=1,,N},特征提取模型 f(x),类别中心向量 WRC×D,超参数sm
      输出:AAML损失
      fori=1toN
       通过网络模型提取到的图像特征 f(xi)D
       计算提取到的图像特征和正负类的类别中心向量夹角 θj,i=1,,C
       对样本所属的类别夹角添加角度间隔 θyiθyi+m
       对样本特征进行缩放 xi=s
       通过Softmax交叉熵损失计算损失
      end
      计算样本损失的均值 L=1Ni=1NLi
    • Table 2. Bands selection

      View table
      View in Article

      Table 2. Bands selection

      算法2高光谱图像波段选择
      Step1:计算高光谱图像各波段的图像熵并进行从大到小排列,从25个波段中筛选出信息量较大的前C = 16个波段
      Step2:在所选出的波段中求两两波段之间的相关系数,分别选出两组相关系数小的波段
      Step3:输入:波段排序后的高光谱图像波段合集 B={Bi}i=1C,初始化波段子集 B1={B1*}B2={B2*},选择两路输入的高光谱波段子集波段数 n=3
      输出:模型两路输入的高光谱波段子集 B1B2
      fori=2 toC
      计算 BB1中各波段与 B1的OIF指数以及 BB2中各波段与 B2的OIF指数
      end
      分别选出使得OIF指数最大的波段加入到各波段子集中
    • Table 3. Ablation experiment

      View table
      View in Article

      Table 3. Ablation experiment

      组合精确率成功率
      MANet0.6890.338
      Ours-AAML(without margin)0.6810.342
      Ours-AAML0.6890.348
      Ours-Attention0.6990.357
      Ours0.7020.365
    Tools

    Get Citation

    Copy Citation Text

    Zonghan SHI, Haitao ZHAO. Hyperspectral target tracking based on attention mechanism and additive angular margin loss[J]. Journal of Applied Optics, 2022, 43(5): 893

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: OE INFORMATION ACQUISITION AND PROCESSING

    Received: Dec. 7, 2021

    Accepted: --

    Published Online: Oct. 12, 2022

    The Author Email: Haitao ZHAO (haitaozhao@ecust.edu.cn)

    DOI:10.5768/JAO202243.0502003

    Topics