• AEROSPACE SHANGHAI
  • Vol. 42, Issue 2, 186 (2025)
Siyu QI, Huijie ZHAO, Hongzhi JIANG, Xudong LI*..., Sihang WANG and Qi GUO|Show fewer author(s)
Author Affiliations
  • Institute of Artificial Intelligence,Beihang University,Beijing100191,China
  • show less
    DOI: 10.19328/j.cnki.2096-8655.2025.02.018 Cite this Article
    Siyu QI, Huijie ZHAO, Hongzhi JIANG, Xudong LI, Sihang WANG, Qi GUO. CNN-LSTM Based Space Object Recognition Method for Sequence Images[J]. AEROSPACE SHANGHAI, 2025, 42(2): 186 Copy Citation Text show less
    Architecture of a space object recognition network for sequential images
    Fig. 1. Architecture of a space object recognition network for sequential images
    Schematic diagram of the single frame feature extraction module design
    Fig. 2. Schematic diagram of the single frame feature extraction module design
    Architecture of an LSTM unit
    Fig. 3. Architecture of an LSTM unit
    Schematic diagram of sampling the experimental sample sequence
    Fig. 4. Schematic diagram of sampling the experimental sample sequence
    Confusion matrix of the validation set of the Top1 model from the noise experiment
    Fig. 5. Confusion matrix of the validation set of the Top1 model from the noise experiment
    Target image with Label 0
    Fig. 6. Target image with Label 0
    Target image with Label 1
    Fig. 7. Target image with Label 1
    Confusion matrix of the validation set of the Top1 model from the attitude experiment
    Fig. 8. Confusion matrix of the validation set of the Top1 model from the attitude experiment
    Comparison of the activation maps for different feature extraction modules
    Fig. 9. Comparison of the activation maps for different feature extraction modules
    训练轮数学习率权重衰减动量
    200.000 34×10-60.9
    Table 1. Parameters for the network training
    实验方法Top1准确率/%mAP准确率/%
    训练集测试集训练集测试集
    所提方法99.6399.4499.7799.48
    VGG[31]+LSTM99.3599.2799.0599.03
    ResNet34[34]+LSTM98.1083.9293.1565.52
    EfficientNet-b0[35]+LSTM95.4294.1794.6393.26
    ViT16[36]+LSTM96.0294.5195.0293.47
    AlexNet[30]+Attention96.7599.3074.9974.58
    Table 2. Comparison of the accuracy of noise experiment
    实验方法Top1准确率/%mAP准确率/%
    训练集测试集训练集测试集
    所提方法96.1994.2794.3693.53
    VGG[31]+LSTM91.2991.1390.5490.47
    ResNet34[34]+LSTM91.0556.6290.5755.19
    EfficientNet-b0[35]+LSTM92.3391.7391.0790.88
    ViT16[36]+LSTM93.1492.2792.2191.07
    AlexNet[30]+Attention88.9583.9487.7381.76
    Table 3. Comparison of the accuracy of randomly sampling attitudes
    实验方法识别时间/s参数量(Flops)/M
    所提方法0.024283.00
    VGG[31]+LSTM0.0476 011.30
    ResNet34[34]+LSTM0.0701 480.20
    EfficientNet-b0[35]+LSTM0.032435.16
    ViT16[36]+LSTM0.09917 630.20
    AlexNet[30]+Attention0.0422 731.60
    Table 4. Comparisons of the recognition time and model size
    Siyu QI, Huijie ZHAO, Hongzhi JIANG, Xudong LI, Sihang WANG, Qi GUO. CNN-LSTM Based Space Object Recognition Method for Sequence Images[J]. AEROSPACE SHANGHAI, 2025, 42(2): 186
    Download Citation